Paper deep dive
AtlasNLP: A Country-Aware Atlas of Dataset Representation in NLP
Joan Nwatu, Tsedeniya Solomon Amare, Longju Bai, Bontu Fufa Balcha, Zayd Bashir, Angana Borah, Zara Burzo, Yubin Choi, Naihao Deng, Samika Gupta, Michel Faloughi, Claude Kwizera, Ziqiao Ma, Cynthia Yacel Fuertes Panizo, Ellie Seehorn, Hui Shen, Jiayi Tang, Zesen Zhao, Boyuan Zheng, Rada Mihalcea
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 9/15/2026, 7:59:47 AM
Summary
The paper introduces AtlasNLP, a country-aware atlas containing over 13,000 NLP dataset records that tracks both the countries represented in datasets and the countries where they are produced. The resource consists of AtlasNLP-Core (automatically extracted from ACL Anthology) and AtlasNLP-Gold (human-curated). The authors demonstrate that dataset coverage is highly uneven, production and representation are geographically asymmetric, and language coverage does not imply geographic representation.
Entities (10)
Relation Signals (10)
AtlasNLP â hascomponent â AtlasNLP-Core
confidence 98% ¡ AtlasNLP includes AtlasNLP-Gold, a human-curated reference set, and AtlasNLP-Core, an ACL-derived large-scale collection.
AtlasNLP â hascomponent â AtlasNLP-Gold
confidence 98% ¡ AtlasNLP includes AtlasNLP-Gold, a human-curated reference set, and AtlasNLP-Core, an ACL-derived large-scale collection.
AtlasNLP-Core â containsrecords â 13,462
confidence 95% ¡ The resulting AtlasNLP-Core contains 13,462 primary dataset-contribution records.
AtlasNLP-Gold â containsrecords â 1,480
confidence 95% ¡ The resulting collection, AtlasNLP-Gold, contains 1,480 human-curated entries
AtlasNLP-Core â derivedfrom â ACL Anthology
confidence 95% ¡ AtlasNLP-Core, an ACL-derived large-scale collection constructed through automated extraction.
AtlasNLP â tracks â producer_countries
confidence 95% ¡ AtlasNLP... tracking both the populations represented and where datasets are produced.
AtlasNLP â tracks â represented_countries
confidence 95% ¡ AtlasNLP... tracking both the populations represented and where datasets are produced.
India â hashighcoverage â AtlasNLP
confidence 90% ¡ The United States, China, India, the United Kingdom, and Germany account for a large share (36.4%) of these associations.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Understanding which countries are represented in NLP datasets is essential for identifying gaps, targeting data collection, measuring progress, and informing AI policy. However, geographic metadata is very rarely available, and country-level representation is often hidden behind broad language-level claims. We introduce AtlasNLP, a country-aware atlas of over 13,000 NLP dataset records across normalized NLP task categories, tracking both the populations represented and where datasets are produced. AtlasNLP includes AtlasNLP-Gold, a human-curated reference set, and AtlasNLP-Core, an ACL-derived large-scale collection. Using this resource, we show that (1) dataset coverage is highly uneven across countries and tasks; (2) dataset production and representation are geographically asymmetric; and (3) language coverage does not imply geographic representation. These findings reveal blind spots in current dataset documentation practices and motivate more explicit geographic metadata for country-aware NLP evaluation.
Tags
Links
- Source: https://arxiv.org/abs/2608.30107v1
- Canonical: https://arxiv.org/abs/2608.30107v1
Trouble viewing inline? Open PDF directly â
Full Text
116,781 characters extracted from source content.
Expand or collapse full text
AtlasNLP: A Country-Aware Atlas of Dataset Representation in NLP Joan Nwatu 1 , Tsedeniya Solomon Amare 1 , Longju Bai 1 , Bontu Fufa Balcha 2 , Zayd Bashir 3 , Angana Borah 1 , Zara Burzo 4 , Yubin Choi 1 , Naihao Deng 1 , Samika Gupta 1 , Michel Faloughi 5 , Claude Kwizera 6 , Ziqiao Ma 1 , Cynthia Yacel Fuertes Panizo 1 , Ellie Seehorn 1 , Hui Shen 1 , Jiayi Tang 1 , Zesen Zhao 1 , Boyuan Zheng 1 , Rada Mihalcea 1 1 University of Michigan, USA, 2 Addis Ababa University, Ethiopia, 3 San Jose State University, USA, 4 Skyline High School, USA, 5 University of Pennsylvania, USA, 6 Carnegie Mellon University, Rwanda jnwatu, mihalcea@umich.edu Abstract Understanding which countries are represented in NLP datasets is essential for identifying gaps, targeting data collection, measuring progress, and informing AI policy. However, geographic metadata is very rarely available, and country- level representation is often hidden behind broad language-level claims. We introduce At- lasNLP, a country-aware atlas of over 13,000 NLP dataset records across normalized NLP task categories, tracking both the populations represented and where datasets are produced. AtlasNLP includes AtlasNLP-Gold, a human- curated reference set, and AtlasNLP-Core, an ACL-derived large-scale collection. Using this resource, we show that (1) dataset coverage is highly uneven across countries and tasks; (2) dataset production and representation are geographically asymmetric; and (3) language coverage does not imply geographic represen- tation. These findings reveal blind spots in cur- rent dataset documentation practices and mo- tivate more explicit geographic metadata for country-aware NLP evaluation. We release At- lasNLP at the AtlasNLP project site. 1 Introduction Efforts to improve representation in NLP increas- ingly emphasize broader dataset coverage across languages, regions, and populations (Burchell, 2024; Bender and Friedman, 2018). Yet it remains difficult to determine which countries and popu- lations are represented, where gaps remain, and whether coverage is improving (Yu et al., 2022; Ranathunga and De Silva, 2022). Most dataset col- lections are organized by language, task, or bench- mark, providing limited visibility into the geogra- phy of dataset representation. This lack of visibility matters because dataset coverage is not globally uniform. As Figure 1 NLP dataset record count per country 01â56â1011â2526â100101â250251â500500+ Highest coverage United States: 631 Sparse coverage 39 countries have 0 82 countries have 1-10 112 countries have 1-25 Figure 1: Global distribution of NLP dataset coverage in AtlasNLP-Core. Countries are bucketed by represented dataset record count, showing a highly uneven distri- bution with a small number of high-coverage countries and many sparsely represented regions. shows, a small number of countries account for a large share of represented dataset records, while many others remain sparsely represented or absent. Coverage also varies across tasks: a country may have datasets for machine translation or sentiment analysis, but little or no coverage for question an- swering, reasoning, safety, or multimodal under- standing. These gaps limit our ability to evaluate NLP systems consistently across countries and pop- ulations and to identify where additional dataset development is most needed (Alabi et al., 2025). Importantly, this also makes it difficult for local governments to plan and assess AI initiatives, since they cannot easily determine what resources are available for their countries or where new invest- ment is most needed. A central reason these patterns are difficult to measure is that geographic metadata is rarely doc- umented explicitly (Faisal et al., 2022). Dataset arXiv:2608.30107v1 [cs.CL] 31 Aug 2026 documentation commonly records language and task, but often omits which countries or popula- tions are represented (Bender and Friedman, 2018; Faisal et al., 2022). Language metadata provides an important but incomplete signal: many languages span multiple countries, and many countries con- tain multiple linguistic communities (Aji et al., 2022; Hershcovich et al., 2022; Tonneau et al., 2024). For example, Spanish is spoken across Spain and much of Latin America, yet datasets labeled as Spanish are often representative only for a subset of these countries. As a result, language- level views can make geographic representation appear broader than the underlying country-level evidence supports, especially when multilingual datasets are treated as evidence of broad popula- tion coverage. To address these limitations, we introduce At- lasNLP, a country-aware atlas of over 13,000 NLP dataset records covering 31 normalized NLP task categories such as machine translation and ques- tion answering. Alongside task and language meta- data, AtlasNLP records both represented coun- tries, i.e., the countries or populations represented in a dataset; and producer countries, i.e., the in- stitutional locations where datasets are produced. The resource consists of two complementary com- ponents: AtlasNLP-Gold, which is a human- curated reference set; and AtlasNLP-Core, an ACL-derived large-scale collection constructed through automated extraction. Using AtlasNLP- Gold, we validate AtlasNLP-Core, and broaden AtlasNLPâs coverage of underrepresented regions. Using AtlasNLP-Core, we analyze patterns of geo- graphic representation, task coverage, and dataset production within the ACL literature. Our contributions are as follows: (1) we release AtlasNLP-Core and AtlasNLP-Gold as country- aware resources for studying NLP dataset geogra- phy; (2) we show that dataset coverage is highly uneven across countries and tasks, with 79.2% of country-task pairs containing no dataset records with documented country representation; (3) we distinguish representation from production, show- ing that many countries are represented primarily through datasets produced outside the country; (4) we show that dataset availability is positively asso- ciated with broader research infrastructure, includ- ing national income and university capacity; and (5) we show that language coverage does not im- ply geographic coverage, motivating more explicit country-level dataset documentation. 2 Related Work Geographic Representation of NLP Datasets. NLP datasets, benchmarks, and model ecosystems are geographically uneven (Bender et al., 2021; Blodgett et al., 2020). AI systems often perform better for populations that are better represented in training and evaluation data (Mihalcea et al., 2025; Nwatu et al., 2026), and model behavior can vary across populations even when tasks ap- pear similar (Hershcovich et al., 2022; Naous et al., 2024). These disparities have been documented across modalities, including vision-language mod- els (Shankar et al., 2017; De Vries et al., 2019; Nwatu et al., 2023, 2025) and speech technologies (Elmadany et al., 2025). Prior work has shown that representation is not only a matter of dataset quantity: languages, re- gions, and populations differ in the tasks, resources, and technologies available to them (Joshi et al., 2020; Blasi et al., 2022; Yu et al., 2022). How- ever, most studies examine geographic or popula- tion bias within individual datasets, models, tasks, languages, or regional settings (Hershcovich et al., 2022; Naous et al., 2024; Alabi et al., 2025; Faisal et al., 2022; Yu et al., 2022; Longpre et al., 2025), but do not provide a unified country-level view of which populations are represented in NLP datasets across regions, who produces those datasets, and how task coverage varies geographically.At- lasNLP addresses this gap by enabling systematic analysis of NLP dataset representation, production, and task coverage across countries. Multilingual and Cross-Lingual Evaluation. NLP evaluation is commonly organized around language. Benchmarks such as XTREME (Hu et al., 2020), XGLUE (Liang et al., 2020), and FLORES (Goyal et al., 2022) expanded evaluations beyond English, while multilingual surveys docu- ment persistent disparities across low-resource and non-Western languages (Joshi et al., 2020; Blasi et al., 2022; Ranathunga and De Silva, 2022; Joshi et al., 2025). However, language is an incomplete proxy for geographic and cultural representation because many languages span multiple countries, and many countries contain multiple linguistic com- munities (Hershcovich et al., 2022; Blodgett et al., 2020; Tonneau et al., 2024; Liu et al., 2025; Pawar et al., 2025). AtlasNLP adds country-level structure alongside language metadata, making it possible to study variation in representation, task coverage, and production within and across languages. Dataset Documentation, Platforms, and Meta- data. Dataset documentation work emphasizes recording dataset origins, collection procedures, intended uses, and potential biases (Gebru et al., 2021; Bender and Friedman, 2018; Longpre et al., 2024). Dataset platforms such as Hugging Face and publication repositories like ACL Anthology have also made resources easier to discover and reuse (Lhoest et al., 2021). Yet geographic metadata re- mains underdeveloped: language and task metadata are often available, while country-level informa- tion is frequently incomplete, implicit, or absent. This makes it difficult to distinguish represented geography, in which populations are represented, from producer geography, in which datasets are institutionally created (Santy et al., 2023). At- lasNLP addresses this gap by introducing struc- tured country-aware metadata for over 13,000 NLP dataset records. 3 Methodology Constructing a country-aware view of the NLP dataset ecosystem requires more than extracting existing metadata. Geographic information is of- ten implicit, inconsistently documented, or absent, and language alone is frequently insufficient for de- termining which populations are represented. We develop AtlasNLP through three stages: (i) design- ing a structured annotation framework and human- curated reference collection, (i) scaling the frame- work over ACL Anthology papers, and (i) audit- ing and refining the extracted dataset records to produce a high-confidence dataset collection. 3.1 Annotation Framework and Schema Design We design a structured annotation framework that records: (i) represented countries, (i) producer countries, (i) task category, (iv) language, (v) modality and licensing, and (vi) attribution method. Task categories are derived from recent ACL and EMNLP thematic areas and normalized into 30 cat- egories. AtlasNLP-Gold additionally includes Raw Corpus for unlabeled or minimally structured text collections. We use a standardized set of 197 geopolitical entities, consisting of the 193 UN member states and two UN observer states, together with Taiwan and Kosovo. A central design choice is the distinc- tion between represented country, the country or population represented in a dataset, and producer country, the institutional location of its creators. This allows us to analyze dataset representation separately from dataset production. Represented-country evidence is classified as ex- plicit, inferred, or unattributed. Explicit attribution requires direct evidence connecting dataset con- tent, participants, data sources, or collection to a country. Inferred attribution captures plausible ge- ographic relevance supported by indirect evidence, such as a geographically specific language variety, community, or data source. Language alone is not sufficient for explicit country attribution. Primary analyses use explicit attribution, while key analyses are repeated with explicit+inferred attribution as a sensitivity analysis. 3.2 Human Curation We construct a human-curated collection to ground the annotation framework, improve coverage of underrepresented regions, and provide a reference set for validation. Contributors come from diverse geographic backgrounds, including Nigeria, China, Peru, the United States, Romania, South Korea, and Ethiopia. Contributors start by selecting the countries they are familiar with and collect associated datasets from ACL Anthology, Hugging Face, Papers With Code, GitHub repositories, institutional archives, and regional dataset collections. To encourage bal- anced coverage, countries are grouped into high-, medium-, and low-resource categories, and con- tributors select country sets spanning multiple re- gions. Contributors are encouraged to identify up to five datasets per country, with particular atten- tion to countries with limited NLP data represen- tation. Each dataset is annotated using the shared schema, including dataset name, task category, lan- guage, modality, licensing, country attribution, and provenance. The curation process has two stages. In the com- pilation stage, contributors independently collect and annotate datasets for their assigned countries. In the cross-validation stage, assignments are shuf- fled and redistributed for secondary review. Re- viewers verify country attribution, task categoriza- tion, metadata consistency, duplicates, and prove- nance, reducing annotator bias and improving con- sistency. The resulting collection, AtlasNLP-Gold, contains 1,480 human-curated entries and serves as a high-quality reference set for evaluating auto- mated extraction. Additional details on Gold con- struction and review are provided in Appendix D. 3.3 Automated Dataset Expansion To scale beyond manual coverage, we apply the annotation framework to ACL Anthology 1 publica- tions. We begin with 119,963 ACL records span- ning 1952â2025. Papers with abstracts are evalu- ated using ModernBERT-base-NLI (Sileo, 2024) against the hypothesis âThis paper proposes a dataset.â We retain papers with entailment scores greater than 0.5, yielding 20,277 candidate dataset papers. Candidate papers are processed using a staged extraction pipeline combining rule-based dataset- role checks with schema-constrained GPT-4o-mini extraction from full paper text. The pipeline ex- tracts dataset identity, task, language, licensing, dataset role, and geographic evidence, together with page- and quote-level provenance. This pro- duces 18,035 successful initial paper-level extrac- tions. A subsequent 400-record human audit reveals recurring failure modes, particularly reuse of ex- isting datasets and country assignments based on affiliation, language, or incidental mentions rather than dataset provenance. We use these findings to tighten the dataset-role and country-attribution criteria and re-evaluate the full extracted collec- tion. Dataset records are retained when the pa- per introduces a dataset, materially extends an ex- isting dataset, or constructs a new dataset com- pilation as a primary contribution. The result- ing AtlasNLP-Core contains 13,462 primary dataset-contribution records. Full extraction, exclusion, and audit details are reported in Ap- pendix A.7. 3.4 Validation and Quality Control We evaluate the final pipeline using three comple- mentary human-reviewed checks. First, we compare post-audit AtlasNLP-Gold against the final AtlasNLP-Core. Matching by ACL paper and dataset identity yields 225 high- confidence aligned dataset records.Among matched Gold dataset records for which AtlasNLP- Core assigns an explicit country, 97.7% share at least one country with the Gold reference set, and 94.7% of individual country assignments made by Core are supported by Gold. Exact task agreement is 79.9%. When inferred country evidence is in- cluded on both Gold and Core, country-assignment precision remains similar at 94.0%. 1 https://aclanthology.org/ Second, the 400-record audit is used to verify that the failure modes identified in the initial ex- traction are substantially reduced after revision. Of 187 dataset records whose country attribution was judged incorrect, 177 (94.7%) no longer remain in the explicit geographic layer. Among the 98 eli- gible cases for which the human review finds no supported country, 97 are correctly left unattributed. Detailed audit results and recall-oriented analyses are reported in Appendix A.7. Finally, because producer geography is derived from author affiliations and is conceptually dis- tinct from represented-country attribution, we eval- uate it independently on 100 human-labeled dataset records.Producer-country extraction achieves 96.9% precision and 97.7% recall at the country- association level, with the primary producer coun- try recovered in all audited cases. 4 AtlasNLP Overview AtlasNLP has two complementary components. AtlasNLP-Core contains 13,462 paper-level dataset records corresponding to primary dataset contributions in ACL Anthology publications. We use the term dataset record rather than unique dataset because related papers may introduce, ex- tend, or compile versions of the same underlying resource, and resolving these relationships into unique dataset entities is not always straightfor- ward. AtlasNLP-Gold contains 1,480 human- curated entries, corresponding to 989 normalized dataset-name groups. We keep the collections separate because they differ in source coverage and metadata completeness: Core provides con- sistent large-scale metadata, including producer- country information from author affiliations, while Gold supplements the ACL-derived collection with human-curated resources, particularly for underrep- resented regions. Table 1 summarizes the final AtlasNLP resource. Represented-country evidence is available at two confidence levels. Explicit country attribution is available for 2,447 Core dataset records (18.2%), while including inferred attribution expands cov- erage to 3,506 dataset records (26.0%). Dataset records without sufficiently supported country ev- idence remain in Core and contribute to analyses that do not require represented-country metadata. Unless otherwise stated, analyses use AtlasNLP- Core. Figure 2: Heatmap of dataset coverage across countries and tasks. Coverage is highly fragmented, with most country-task pairs lacking represented dataset records and even high-coverage countries concentrated in a subset of tasks. Only a subset of country labels is shown for readability, with labels placed at regular intervals (approximately every fifth country). Best viewed in color. 5 Dataset Geography Analysis 5.1Dataset Coverage is Highly Uneven Across Countries and Tasks We first examine which countries are represented in NLP datasets using explicit represented-country attribution. Coverage follows a pronounced long- tail distribution. Across AtlasNLP-Core, 2,447 dataset records have explicit country attribution, corresponding to 4,421 country-record associations across 158 of 197 countries. The United States, China, India, the United Kingdom, and Germany account for a large share (36.4%) of these associa- tions. At the other end of the distribution, 39 coun- tries have no explicitly attributed dataset records, and 121 of 197 countries have ten or fewer repre- sented dataset records. Coverage is even sparser across countries and tasks. As shown in Figure 2, nearly four in five country-task pairs have no explicitly attributed dataset records (79.2%). This pattern remains under explicit+inferred attribution, where three- quarters of pairs are still empty (75.0%). Geo- graphic imbalance therefore reflects not only how many represented dataset records each country has, but also which NLP tasks they cover. Country attribution is also systematically miss- ing. Among unattributed Core records, 70.7% in- clude English, compared with 49.9% of explicitly attributed dataset records. Because English is glob- ally distributed, language alone provides little evi- dence of represented country. We therefore inter- pret low or zero coverage as gaps in documented country representation within ACL dataset contri- butions, not evidence that no NLP resources exist. StatisticValue AtlasNLP-Core dataset records13,462 AtlasNLP-Gold entries / normalized name groups1,480 / 989 Core dataset records w/explicit country attribution2,447 (18.2%) Core dataset records w/explicit+inferred attribution3,506 (26.0%) Core countries represented (explicit / +inferred)158 / 168 Gold countries represented (explicit / +inferred)192 / 197 Core country-record associations (explicit / +inferred)4,421 / 6,001 Task categories (Core / AtlasNLP overall)30 / 31 Audited languages (Core / AtlasNLP overall)1,239 / 1,245 Multilingual dataset records (Core)3,063 Producer-country metadata (Core)13,224 (98.2%) Table 1: Summary statistics for AtlasNLP. AtlasNLP- Core contains vetted primary dataset-contribution records. Represented-country statistics distinguish ex- plicit attribution from the broader explicit+inferred sen- sitivity layer. 5.2 Dataset Production and Representation are Asymmetric Dataset coverage alone does not show who pro- duces the datasets used to represent different coun- tries. We therefore compare represented country with producer country, derived from the institu- tional locations of dataset creators based on author affiliations. Producer geography captures institu- tional location rather than researcher identity, re- flecting differences in research capacity, agendas, and environments. To construct the production-representation analy- sis, we first restrict the explicit represented-country layer to records for which producer-country meta- data is available.This reduces the 2,447 ex- plicitly attributed Core records (4,421 country- record associations) to 2,396 records and 4,305 represented-country associations. Records with- out recovered producer geography remain in the coverage and task analyses but cannot contribute to production-representation ratios. For example, BizBench (Krumdick et al., 2024) and the Online Tenant Reviews dataset (Haber and Waks, 2021) have explicit U.S. representation in AtlasNLP but no recovered producer country. We then expand each retained represented- country association over all producer countries as- sociated with that record. Thus, a dataset record representingRcountries and associated with insti- tutions inPproducer countries contributesRĂ P producer-representation associations. This expan- sion yields 6,533 producer-representation associa- tions from the 2,396 eligible records and preserves multi-country representation and cross-country col- laboration rather than forcing either side to a single country. Figure 3 compares two complementary mea- sures. The x-axis shows content self-representation: among represented records about a country with producer metadata, the share that include that same country among their producer countries. The y- axis shows producer self-representation: among expanded producer-representation associations in- volving a producer country, the share whose repre- sented country is that same country. Together, these measures distinguish countries whose representa- tion is primarily produced locally or externally, and producers that focus primarily on themselves or on other countries. 0.000.250.500.751.00 Content self-representation (share of dataset records about country X produced domestically) 0.0 0.2 0.4 0.6 0.8 1.0 Producer self-representation (share of producer representation associations involving a country that represent that same country) Producer vs. content self-representation explicit attribution Regime Locally produced, self-focused Locally produced, outward-facing Externally represented, self-focused Externally represented, outward-facing Hollow = producer_count < 10 Brazil China Denmark Egypt India Nigeria Russia United States Figure 3: Dataset production and representation by country under explicit attribution. The x-axis mea- sures the share of datasets about a country produced domestically; the y-axis measures the share of producer- representation associations involving a country that represent itself. Dashed lines define four production- representation regimes. Point size reflects dataset vol- ume; hollow points indicate countries with fewer than 10 producer associations. The resulting patterns are strongly asymmet- ric. Among countries with at least 10 represented records with producer metadata and 10 producer- representation associations, 39 of 62 (62.9%) have content self-representation below 0.5, meaning that most dataset records representing them were pro- duced by institutions outside the country. This pattern is similar without the minimum-count re- striction and under explicit+inferred attribution. Figure 3 shows the broader distribution, while Table 2 reports values for representative countries in each regime. China and India are both locally produced and self-focused, while the United States and Denmark have high domestic representation but produce many datasets about other countries. Nigeria and Egypt are more externally represented and outward-facing. Russia and Brazil show a dif- ferent pattern: much of their representation is pro- duced externally, while their own dataset produc- Country Content self-rep. Producer self-rep. Represented records w/prod. Producer-rep associations United States0.5970.2656201394 China0.7470.615364442 India0.5370.513227238 Denmark0.7860.2782879 Nigeria0.2160.2763729 Egypt0.1090.1155552 Russia0.1400.7005010 Brazil0.3860.6007045 Table 2: Examples of production-representation asym- metry under explicit country attribution. Represented- record counts are restricted to records with producer metadata; producer-side counts are expanded producer- representation associations. All highlighted countries have at least 10 represented records with producer meta- data and 10 producerârepresentation associations. tion is comparatively self-focused. Production is also geographically concen- trated.The United States accounts for 21.3% of ex- panded producer-representation associations, while the ten largest producer countries account for 61% of these associations. This concentration persists under explicit+inferred attribution. Together, these results show that geographic representation in NLP depends not only on which countries appear in datasets, but also on where the institutions produc- ing those datasets are located (Santy et al., 2023). 5.3 Dataset Availability Tracks Research Infrastructure We next examine whether geographic dataset record coverage aligns with broader research infras- tructure. Using World Bank income categories and national university counts, we compare represented- country coverage with two coarse indicators of in- stitutional capacity. Figure 4(A) shows a strong income gradient: the median number of represented dataset records is substantially higher for high-income countries than for the other income groups. Figure 4(B) similarly shows a positive associa- tion between national university counts and dataset coverage (Pearsonr = .65; SpearmanĎ = .61). The relationship strengthens slightly under ex- plicit+inferred attribution (r = .70; Ď = .67). These patterns show that dataset availability is positively associated with broader research infras- tructure. The relationship is not deterministic, how- ever: countries with similar levels of institutional capacity can still differ substantially in coverage, reflecting the roles of language, publication ecosys- tems, and regional research priorities. Low Lower-mid Upper-mid High 10 0 10 1 10 2 Represented dataset records (log scale) (A) Coverage by income group | explicit 10 1 10 0 10 1 10 2 10 3 # universities (log scale) 10 1 10 0 10 1 10 2 10 3 Represented dataset records (log scale) United States India Nigeria (B) Coverage vs. university count | explicit Income Low Lower-mid Upper-mid High Figure 4: Relationship between represented-country dataset records and research infrastructure under ex- plicit attribution. (A) Coverage by World Bank income group. (B) Coverage increases with national univer- sity counts on log-transformed values. Color indicates income group. Best viewed in color. 5.4 Task Coverage is Fragmented Across Countries Geographic disparities extend beyond dataset avail- ability to the range of NLP tasks represented. We measure each countryâs task breadth, the number of task categories represented, and top-3 task share, the proportion of represented dataset records con- centrated in its three most common tasks. We restrict the analysis to countries with at least 10 represented dataset records and classify portfolios with a top-3 share above 0.60 as specialist, and the remainder as generalist. Even among these better-represented countries, task coverage remains uneven. Under explicit at- tribution, the median country spans 11 of 30 task categories, while its three most common tasks ac- count for 55.6% of represented dataset records. Ex- plicit+inferred attribution increases median breadth to 13 tasks, while concentration remains similar at 54.3%. Table 3 illustrates this variation. Countries such as the United States, China, India, and France have broad generalist portfolios, whereas Belgium, Country Represented records Task breadth Top-3 share Portfolio United States631270.50Generalist China368270.48Generalist India233220.46Generalist France144220.44Generalist Egypt58150.53Generalist Mexico49140.61Specialist Belgium2670.81Specialist Bahrain1570.73Specialist Israel2080.70Specialist Oman2090.70Specialist Table 3: Examples of country-level task portfolios under explicit attribution. Task breadth is the number of repre- sented task categories, and top-3 share is the proportion of a countryâs represented dataset records concentrated in the three most common tasks. We label portfolios with top-3 share above 0.60 as specialist and the re- mainder as generalist. Countries with fewer than 10 represented dataset records are excluded. Bahrain, Israel, and Oman are concentrated in a much narrower set of tasks. Thus, geographic cov- erage does not imply broad task coverage: a coun- try may be represented in NLP datasets but remain concentrated in only a small subset of tasks. 5.5 Language Metadata Does Not Imply Geographic Coverage A central challenge in analyzing dataset geography is that represented-country information is often un- available or insufficiently documented. Datasets more commonly report language, making language an appealing proxy for geographic coverage even though the same language may span many coun- tries and populations. AtlasNLP addresses this gap by distinguishing explicit represented-country attribution from addi- tional inferred attribution. Inferred attribution was intentionally conservative, as language alone was not used to assign broadly distributed languages such as English, French, Spanish, Portuguese, Ara- bic, or Swahili to countries; such assignments re- quired additional country-specific contextual ev- idence. Even under explicit+inferred attribution, however, a represented country could not be es- tablished for 9,956 of 13,462 Core dataset records (74.0%). This substantial attribution gap illustrates why language metadata alone cannot resolve the geography of much of the NLP dataset ecosystem. Figure 5 shows substantial variation in coun- try concentration across widely used languages. Chinese-language records are overwhelmingly as- sociated with China (89.2%). For Portuguese, Brazil and Portugal together account for about 0.00.20.40.60.81.0 Share of dataset records within language Chinese Portuguese French English Spanish Language Country concentration within language-labeled dataset records China Taiwan United States Brazil Portugal Germany France Switzerland Canada United Kingdom India Spain Mexico Peru Others Figure 5: Country concentration within language- labeled dataset records under explicit attribution. Coun- try representation varies substantially within widely used languages: some are dominated by one or a few countries, while others are more geographically dis- tributed. Language coverage therefore does not directly indicate which countries are represented. 69% of represented dataset records, while France, Switzerland, and Canada account for about 63% of French records. English is more geographically distributed, yet the United States, United Kingdom, and India still account for about half of represented dataset records. These patterns show why language alone is an incomplete measure of geographic representation. A language may span many countries without those countries being represented equally, or at all, within NLP datasets. Explicit country-aware metadata is therefore necessary for observing population-level gaps that language metadata alone can obscure. 6 Lessons Learned and Implications Document represented populations explicitly. A major challenge in constructing AtlasNLP was the absence of recoverable country-level prove- nance. Even with explicit+inferred attribution, a represented country could not be established for 74% of AtlasNLP-Core dataset records. This does not imply that these datasets lack geographic con- text, but that it is often not documented in a form that can be reliably recovered. Dataset documenta- tion should therefore include represented countries or populations, geographic scope, and data prove- nance to support more transparent and accurate coverage claims (Longpre et al., 2024). Distinguish explicit from inferred geography. Plausible geographic signals should not be treated as equivalent to documented provenance. Separat- ing explicit and inferred attribution makes uncer- tainty visible and allows geographic analyses to be evaluated under both strict and broader assump- tions. Treat language as necessary but insufficient. Language remains essential for organizing NLP re- sources, but it should not be treated as a proxy for geographic representation (Tonneau et al., 2024). Countries sharing a language can have very differ- ent levels of representation, and language-labeled resources may remain concentrated in only a subset of the populations that use that language. Separate representation from production. Dataset geography is shaped by both who is repre- sented and who produces the data. Distinguishing represented geography from producer geography makes institutional asymmetries in dataset creation more visible. Track task coverage, not only dataset counts. Countries with some dataset coverage may still lack coverage across large parts of the NLP task space. Reporting task-level geographic coverage can help identify where dataset development is broad, nar- row, or missing. Make geographic metadata first-class dataset infrastructure. Large dataset platforms such as Hugging Face have improved resource visibility, but population-level coverage remains difficult to audit without standardized metadata for repre- sented populations, geographic provenance, and producer geography. AtlasNLP offers one schema for making these distinctions explicit and geo- graphic gaps easier to identify. 7 Conclusion We introduced AtlasNLP, a country-aware atlas of NLP dataset contributions that supports analysis of geographic representation, dataset production, and task coverage. By distinguishing the popula- tions represented in datasets from the institutional locations where datasets are produced, and explicit from inferred geographic attribution, AtlasNLP makes visible patterns that are difficult to observe in language-centered resource collections. Our analysis shows that NLP dataset coverage is highly uneven across countries and tasks, pro- duction is geographically concentrated and asym- metric, and language coverage does not imply ge- ographic coverage. These findings highlight limi- tations in current dataset documentation practices and show why geographic diversity cannot be as- sessed reliably from language metadata or producer geography alone. We release AtlasNLP at the At- lasNLP project site and hope it supports future work on dataset transparency, geographic repre- sentation, and population-aware evaluation, while encouraging geographic metadata to become a first- class component of dataset documentation. Limitations AtlasNLP is subject to limitations in metadata qual- ity, coverage, and scope. First, represented-country attribution remains incomplete and non-random. Our precision-first strategy leaves dataset records unattributed when country provenance cannot be established confidently, particularly for globally distributed languages. Consequently, low or zero country coverage should be interpreted as limited documented representation within the ACL liter- ature, rather than evidence that no relevant NLP resources exist. Second, AtlasNLP-Core consists of paper-level dataset-contribution records rather than fully dedu- plicated dataset entities. Related dataset records may describe extensions, compilations, or versions of the same underlying resource, for which entity- level deduplication is not always straightforward. AtlasNLP therefore measures documented dataset activity, not the number of unique datasets, dataset size, quality, or downstream evaluation capacity. Multi-country dataset records may also contribute to several country totals. These counts should there- fore be interpreted as indicators of research visibil- ity and coverage rather than data volume. Third, most Core annotations are automatically extracted. Validation against AtlasNLP-Gold and targeted audits indicate high precision for the ge- ographic layers used in our analyses, but errors may remain, particularly for complex multi-country scope. Gold also shares the AtlasNLP annotation schema and is not a fully independent ground truth. Fourth, the ModernBERT NLI filter used to iden- tify candidate dataset papers was not evaluated for population-wide recall across the full ACL Anthol- ogy. AtlasNLP should therefore not be interpreted as an exhaustive census of all dataset contributions in ACL, and we avoid temporal claims that would depend on uniform recall across publication peri- ods. Fifth, producer geography is derived from au- thor affiliations and represents institutional location rather than researcher identity, nationality, or intent. This distinction is especially important for diaspora researchers and cross-border collaborations. Finally, AtlasNLP-Core is intentionally scoped to ACL Anthology publications.Datasets re- leased only through Hugging Face, GitHub, in- dustry repositories, government portals, commu- nity archives, or non-ACL venues are therefore underrepresented. Our normalized task schema also trades granularity for cross-country compara- bility and may obscure locally specific or niche task distinctions. Ethical Considerations AtlasNLP makes country-level gaps in NLP dataset representation more visible, but country-level anal- ysis can also be overinterpreted. Countries are not equivalent to cultures, languages, ethnic groups, or lived experiences, and substantial variation exists within national boundaries. We therefore caution against using AtlasNLP to rank countries, essential- ize populations, or treat country labels as complete representations of identity or culture. Geographic attribution also carries uncertainty. Inferred country assignments capture plausible ge- ographic relevance under deliberately conservative rules; they should not be interpreted as definitive claims about the identity of dataset contributors or populations. For this reason, AtlasNLP preserves the distinction between explicit and inferred attri- bution and uses explicit provenance for its primary geographic analyses. There is also a risk that evidence of underrepre- sentation could be used to normalize poorer model performance for some regions rather than moti- vate its improvement. Our goal is the opposite: to identify gaps in representation, production, and task coverage so that future data collection and evaluation can be more inclusive and accountable. Likewise, absence or low coverage in AtlasNLP should not be interpreted as evidence that a country or community lacks NLP resources. Producer-country metadata should be interpreted carefully. It reflects institutional affiliation, not researcher nationality, identity, or intent. Externally produced datasets may result from collaboration, diaspora scholarship, or resource sharing; our aim is not to privilege domestic production, but to make production and representation patterns visible. AtlasNLP-Gold was constructed through a col- laborative research effort rather than paid crowd- sourcing. Contributors compiled and validated en- tries using shared guidelines and were credited as coauthors when they completed the agreed cura- tion and validation workload. Machine-assisted ex- tensions followed the same human curation goals and guidelines. Gold was deliberately curated to broaden geographic coverage and should therefore not be interpreted as a natural sample of the global NLP dataset ecosystem. Acknowledgments We thank the anonymous reviewers for their feed- back, and the members of the Language and In- formation Technologies lab at the University of Michigan for the insightful discussions during the early stage of the project. This project was par- tially funded by a grant from OpenAI and a grant from the Survival and Flourishing Fund and a grant from the University of Michigan under the Strate- gic Initiative Fund. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not neces- sarily reflect the views of Open AI or the Survival and Flourishing Fund or the University of Michi- gan. References Alham Fikri Aji, Genta Indra Winata, Fajri Koto, Samuel Cahyawijaya, Ade Romadhony, Rahmad Ma- hendra, Kemal Kurniawan, David Moeljadi, Radi- tyo Eko Prasojo, Timothy Baldwin, Jey Han Lau, and Sebastian Ruder. 2022. One country, 700+ lan- guages: NLP challenges for underrepresented lan- guages and dialects in Indonesia. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7226â7249, Dublin, Ireland. Association for Computational Linguistics. Jesujoba Oluwadara Alabi, Michael A. Hedderich, David Ifeoluwa Adelani, and Dietrich Klakow. 2025. Charting the landscape of African NLP: Mapping progress and shaping the road ahead. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 27807â27841, Suzhou, China. Association for Computational Lin- guistics. Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587â604. Emily M Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM confer- ence on fairness, accountability, and transparency, pages 610â623. Damian Blasi, Antonios Anastasopoulos, and Gra- ham Neubig. 2022. Systematic inequalities in lan- guage technology performance across the worldâs languages. In Proceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5486â5505. Su Lin Blodgett, Solon Barocas, Hal DaumĂŠ Iii, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of âbiasâ in nlp. In Proceed- ings of the 58th annual meeting of the association for computational linguistics, pages 5454â5476. Laurie Vear Burchell. 2024.Improving Natural Language Processing for Under-Served Languages through Increased Training Data Diversity. Ph.D. thesis, University of Edinburgh. Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. 2019. Does object recog- nition work for everyone? In Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition workshops, pages 52â59. AbdelRahim A. Elmadany,Sang Yun Kwon, Hawau Olamide Toyin, Alcides Alcoba Inciarte, Hanan Aldarmaki, and Muhammad Abdul-Mageed. 2025.Voice of a continent: Mapping Africaâs speech technology frontier. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 11028â11050, Suzhou, China. Association for Computational Linguistics. Fahim Faisal, Yinkai Wang, and Antonios Anastasopou- los. 2022. Dataset geography: Mapping language data to language users. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3381â 3411. Timnit Gebru, Jamie Morgenstern, Briana Vec- chione, Jennifer Wortman Vaughan, Hanna Wallach, Hal DaumĂŠ Iii, and Kate Crawford. 2021. Datasheets for datasets. Communications of the ACM, 64(12):86â 92. Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng- Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Kr- ishnan, MarcâAurelio Ranzato, Francisco GuzmĂĄn, and Angela Fan. 2022. The flores-101 evaluation benchmark for low-resource and multilingual ma- chine translation. Transactions of the Association for Computational Linguistics, 10:522â538. Adam Haber and Zeev Waks. 2021. Classification and geotemporal analysis of quality-of-life issues in ten- ant reviews. In Findings of the Association for Com- putational Linguistics: EMNLP 2021, pages 2541â 2553, Punta Cana, Dominican Republic. Association for Computational Linguistics. Daniel Hershcovich, Stella Frank, Heather Lent, Miryam de Lhoneux, Mostafa Abdou, Stephanie Brandl, Emanuele Bugliarello, Laura Cabello Pi- queras, Ilias Chalkidis, Ruixiang Cui, Constanza Fierro, Katerina Margatina, Phillip Rust, and Anders Søgaard. 2022. Challenges and strategies in cross- cultural NLP. In Proceedings of the 60th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 6997â7013, Dublin, Ireland. Association for Computational Lin- guistics. Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan Firat, and Melvin Johnson. 2020. Xtreme: A massively multilingual multi-task bench- mark for evaluating cross-lingual generalisation. In International conference on machine learning, pages 4411â4421. PMLR. Aditya Joshi, Raj Dabre, Diptesh Kanojia, Zhuang Li, Haolan Zhan, Gholamreza Haffari, and Doris Dip- pold. 2025. Natural language processing for dialects of a language: A survey. ACM Computing Surveys, 57(6):1â37. Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the nlp world. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 6282â6293. Michael Krumdick, Rik Koncel-Kedziorski, Viet Dac Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. 2024. BizBench: A quantitative reasoning benchmark for business and finance. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8309â8332, Bangkok, Thailand. Association for Computational Linguistics. Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Ĺ aĹĄko, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, and 13 others. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Nat- ural Language Processing: System Demonstrations, pages 175â184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. Yaobo Liang, Nan Duan, Yeyun Gong, Ning Wu, Fenfei Guo, Weizhen Qi, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, Xiaodong Fan, Ruofei Zhang, Rahul Agrawal, Edward Cui, Sining Wei, Taroon Bharti, Ying Qiao, Jiun-Hung Chen, Winnie Wu, and 5 others. 2020. XGLUE: A new benchmark dataset for cross-lingual pre-training, understanding and gen- eration. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6008â6018, Online. Association for Computational Linguistics. Chen Cecilia Liu, Iryna Gurevych, and Anna Korho- nen. 2025. Culturally aware and adapted nlp: A taxonomy and a survey of the state of the art. Trans- actions of the Association for Computational Linguis- tics, 13:652â689. Shayne Longpre, Robert Mahari, Naana Obeng-Marnu, William Brannon, Tobin South, Katy Gero, Sandy Pentland, and Jad Kabbara. 2024. Data authenticity, consent, & provenance for ai are all broken: what will it take to fix them? arXiv preprint arXiv:2404.12691. Shayne Longpre, Nikhil Singh, Manuel Cherep, Kusha- gra Tiwary, Joanna Materzynska, William Brannon, Robert Mahari, Naana Obeng-Marnu, Manan Dey, Mohammed Hamdy, Nayan Saxena, Ahmad Mustafa Anis, Emad Alghamdi, Minh Chien Vu, Da Yin, Kun Qian, Yizhi Li, Minnie Liang, An Dinh, and 24 oth- ers. 2025. Bridging the data provenance gap across text, speech, and video. In International Conference on Learning Representations, volume 2025, pages 60592â60670. Rada Mihalcea, Oana Ignat, Longju Bai, Angana Borah, Luis Chiruzzo, Zhijing Jin, Claude Kwizera, Joan Nwatu, Soujanya Poria, and Thamar Solorio. 2025. Why ai is weird and shouldnât be this way: towards ai for everyone, with everyone, by everyone. In Pro- ceedings of the AAAI conference on artificial intelli- gence, volume 39, pages 28657â28670. Tarek Naous, Michael J Ryan, Alan Ritter, and Wei Xu. 2024. Having beer after prayer? measuring cultural bias in large language models. In Proceedings of the 62nd annual meeting of the association for compu- tational linguistics (volume 1: Long papers), pages 16366â16393. Joan Nwatu, Longju Bai, Oana Ignat, and Rada Mi- halcea. 2026. Culture affordance atlas: Reconcil- ing object diversity through functional mapping. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 39089â39097. Joan Nwatu, Oana Ignat, and Rada Mihalcea. 2023. Bridging the digital divide: Performance variation across socio-economic factors in vision-language models. In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 10686â10702. Joan Nwatu, Oana Ignat, and Rada Mihalcea. 2025. Uplifting lower-income data: Strategies for socioeco- nomic perspective shifts in large multi-modal models. In Proceedings of the 2025 Conference of the Na- tions of the Americas Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (Volume 1: Long Papers), pages 2127â2144. Siddhesh Pawar, Junyeong Park, Jiho Jin, Arnav Arora, Junho Myung, Srishti Yadav, Faiz Ghifari Haznitrama, Inhwa Song, Alice Oh, and Isabelle Au- genstein. 2025. Survey of cultural awareness in lan- guage models: Text and beyond. Computational Linguistics, 51(3):907â1004. Surangika Ranathunga and Nisansa De Silva. 2022. Some languages are more equal than others: Probing deeper into the linguistic disparity in the nlp world. In Proceedings of the 2nd Conference of the Asia- Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Confer- ence on Natural Language Processing (Volume 1: Long Papers), pages 823â848. Sebastin Santy, Jenny Liang, Ronan Le Bras, Katharina Reinecke, and Maarten Sap. 2023. NLPositionality: Characterizing design biases of datasets and models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9080â9102, Toronto, Canada. Association for Computational Linguistics. Shreya Shankar, Yoni Halpern, Eric Breck, James At- wood, Jimbo Wilson, and D Sculley. 2017. No classi- fication without representation: Assessing geodiver- sity issues in open data sets for the developing world. arXiv preprint arXiv:1711.08536. Damien Sileo. 2024. tasksource: A large collection of NLP tasks with a structured dataset preprocess- ing framework. In Proceedings of the 2024 Joint International Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 15655â15684, Torino, Italia. ELRA and ICCL. Manuel Tonneau, Diyi Liu, Samuel Fraiberger, Ralph Schroeder, Scott A Hale, and Paul RĂśttger. 2024. From languages to geographies: Towards evaluating cultural bias in hate speech datasets. In Proceedings of the 8th Workshop on Online Abuse and Harms (WOAH 2024), pages 283â311. Xinyan Yu, Trina Chatterjee, Akari Asai, Junjie Hu, and Eunsol Choi. 2022. Beyond counting datasets: A survey of multilingual dataset construction and necessary resources. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 3725â3743. A Pipeline and Construction Details This appendix provides additional details on the construction, audit, and validation of AtlasNLP- Core and AtlasNLP-Gold. Figure 6 summarizes the full pipeline. AtlasNLP-Core is constructed in two stages: an initial large-scale extraction from ACL Anthology papers, followed by audit-driven refinement of dataset eligibility and geographic at- tribution. AtlasNLP-Gold is constructed through human curation, cross-validation, and subsequent audit and refinement. The sections below describe the metadata schema, task taxonomy, country- attribution policy, language normalization, auto- mated extraction, post-extraction audit, producer- country extraction, and construction of the country- level analysis layers. AtlasNLP-GoldAtlasNLP-Core Contributor country assignment 197-entity ontology Dataset search and annotation ACL Anthology, Hugging Face, GitHub, regional sources, etc. Cross-validation country, task, links, metadata, and provenance Machine-assisted extension and post-curation audit AtlasNLP-Gold 1,480 human-curated entries 989 normalized dataset-name groups ACL Anthology bibliography 119,963 records Records with non-empty abstracts 71,847 ModernBERT NLI filter 20,277 candidates entailment score> 0.5 Role triage + GPT-4o-mini extraction 18,035 successful initial extractions 400-record human audit dataset role + represented-country attribution Audit-driven full re-evaluation stricter dataset-role and geographic-attribution criteria AtlasNLP-Core 13,462 primary dataset-contribution records Validation and final quality checks GoldâCore: 225 high-confidence dataset matches¡400-record post-revision audit verification 100-record producer-country audit: precision 96.9%, recall 97.7% Explicit geography 2,447 dataset records 4,421 country- record associations 158 countries Explicit + inferred 3,506 dataset records 6,001 country- record associations 168 countries Producer geography 13,224 dataset records 98.2% coverage Figure 6: Overview of AtlasNLP construction, audit, and validation. AtlasNLP-Core proceeds from ACL candidate selection through initial automated extraction and audit-driven re-evaluation, yielding 13,462 final primary dataset- contribution records. AtlasNLP-Gold contains 1,480 human-curated entries corresponding to 989 normalized dataset-name groups. Final geographic layers distinguish explicit attribution from the broader explicit+inferred sensitivity analysis. The initial ACL pipeline begins with 119,963 ACL Anthology records, of which 71,847 contain non-empty abstracts. ModernBERT-base-NLI as- signs 23,674 papers an entailment label for the hy- pothesis that the paper proposes a dataset; 20,277 also exceed the 0.5 entailment threshold and en- ter the extraction pipeline. Role triage and LLM- assisted metadata extraction produce 18,035 suc- cessful initial paper-level extractions. A subsequent 400-record human audit identifies recurring dataset- role and geographic-attribution errors and moti- vates a full re-evaluation of these records. The final AtlasNLP-Core contains 13,462 primary dataset- contribution records. A.1 Metadata Schema AtlasNLP uses a shared conceptual schema for human-curated and automatically extracted dataset records. The schema captures dataset identity and provenance, dataset role, task and language cov- erage, represented and producer geography, and additional dataset characteristics. The final Core release also retains audit and normalization fields used to document eligibility and geographic attri- bution. Table 4 summarizes the principal metadata fields. The schema separates represented country, which captures the geographic population or con- text represented by a dataset, from producer coun- try, which captures where dataset production is in- stitutionally located. It also separates dataset iden- tity from dataset relationship: related papers may introduce, materially extend, compile, or reuse the same underlying resource. Final AtlasNLP-Core eligibility is therefore defined at the level of the paperâdataset contribution rather than by dataset- name deduplication alone. FieldDescription Dataset nameName of the dataset, benchmark, resource, or dataset compilation. Paper linkPaper URL or citation associated with the dataset contribution. Dataset linkOfficial dataset, benchmark, repository, GitHub, or Hugging Face URL. Dataset relationship / role Relationship between the paper and target dataset: introduced, materially extended, newly compiled, reused from prior work, not identifiable, or ambiguous. This field determines AtlasNLP-Core eligibility. Task categoryNormalized NLP task category from the AtlasNLP taxonomy. Language(s)Dataset languages or language varieties, normalized and audited when possible. Represented countryCountry or population represented by the dataset content, participants, sources, or collection context. Country attributionGeographic evidence layer: explicit, inferred, or unattributed when evidence is insufficient. Producer countryInstitutional location of dataset creators, derived from author affiliations for ACL-derived records. ModalityDataset modality, including text-only and multimodal resources. Multiple-choice formatWhether the dataset or benchmark uses a multiple-choice format. SyntheticWhether content is machine-translated, model-generated, or otherwise synthetic, when recover- able. LicenseDataset license, when available. Provenance / evidenceSource information supporting extracted metadata, including page- and quote-level evidence where available. Audit / normalization fields Eligibility decisions, attribution status, normalized country lists, processing status, and related post-audit quality-control fields. Table 4: Principal metadata fields used in AtlasNLP. The final Core release also retains audit and normalization fields documenting dataset eligibility and geographic-attribution decisions. A.2 Task Taxonomy AtlasNLP uses a normalized task taxonomy de- rived from recent ACL and EMNLP thematic areas. AtlasNLP-Core contains 30 task categories, while AtlasNLP-Gold additionally includes Raw Corpus, yielding 31 categories across AtlasNLP overall. Ta- ble 5 lists the final taxonomy and short labels used in figures. A.3 Country Attribution Rules AtlasNLP uses a standardized set of 197 geopo- litical entities, consisting of the 193 UN member states and two UN observer states, together with Taiwan and Kosovo. Represented-country attribu- tion is separated from producer geography and is classified into three confidence levels: explicit, in- ferred, and unattributed. â˘Explicit attribution. A country is assigned explicitly when dataset documentation di- rectly connects the dataset content, partici- pants, data sources, collection process, tar- get population, or benchmark scope to that country. Examples include data collected in a named country, a benchmark explicitly de- signed for a countryâs population, or source material drawn from a clearly identified na- tional context. â˘Inferred attribution. A country is assigned as inferred when geographic relevance is plau- sible from indirect but geographically infor- mative evidence, such as a country-specific language variety, a locally defined community, or a data source with a sufficiently narrow geo- graphic scope. Inferred attribution is retained separately from explicit attribution and is used only in the broader sensitivity analyses. â˘Unattributed.A dataset record remains unattributed when available evidence is in- sufficient to support either explicit or con- servative inferred geography. Broadly dis- tributed languages, author affiliations, inci- dental country mentions, and the geographic location of the producing institution are not by themselves treated as evidence of represented country. This distinction is important because represented country and producer country capture different con- cepts. Author affiliation may establish producer geography but does not establish which population the dataset represents. The post-extraction audit identified affiliation-based assignment, language- only assignment, and incidental country mentions as recurring sources of erroneous represented- country attribution; the final pipeline therefore ap- plies these exclusions explicitly. For multi-country datasets, all countries sup- Task categoryShort label Code ModelsCode Models Computational Social Science, Cultural Analytics, NLP for Social GoodCSS / NLP4SG Dialogue and Interactive SystemsDialogue Discourse, Pragmatics, and ReasoningDiscourse / Reasoning Efficient / Low-Resource Methods for NLP, LLM EfficiencyLow-resource / Efficiency Ethics, Bias, and FairnessEthics / Bias Generalizability and TransferTransfer Hierarchical Structure PredictionHierarchical Human-AI Interaction / Cooperation, Human-Centered NLPHuman-Centered NLP Information Extraction, Retrieval, and Text MiningIE / Retrieval Interpretability, Model Analysis, Transparency, and ExplainabilityInterpretability Language ModelingLanguage Modeling Linguistic Theories, Cognitive Modeling, and PsycholinguisticsPsycholinguistics Machine Learning for NLPML for NLP Machine TranslationMT Mathematical, Symbolic, and Logical Reasoning in NLPMath / Logic Multilinguality and Language DiversityMultilinguality Multimodality and Language Grounding to Vision, Robotics, and BeyondMultimodality Natural Language GenerationNLG Neurosymbolic Approaches to NLPNeurosymbolic NLP ApplicationsNLP Applications Phonology, Morphology, and Word SegmentationPhonology / Morphology Question AnsweringQA Raw CorpusRaw Corpus Retrieval-Augmented Language ModelsRAG Safety and Alignment in LLMsSafety Semantics: Lexical and Sentence Level, Textual InferenceSemantics / Inference Sentiment Analysis, Stylistic Analysis, and Argument MiningSentiment / Style / Argument Speech Recognition, Text-to-Speech, and Spoken Language UnderstandingSpeech SummarizationSummarization Syntax: Tagging, Chunking, and ParsingSyntax Table 5: AtlasNLP-Core+Gold task taxonomy. AtlasNLP-Core contains 30 ACL/EMNLP-derived categories; AtlasNLP-Gold adds the human-curated Raw Corpus category. Short labels are used in compact visualizations. ported by the available evidence are retained. A sin- gle dataset record can therefore contribute to mul- tiple country-record associations, while remaining one dataset record in analyses that do not expand by represented country. Synthetic datasets follow the same attribution policy. Producer country is derived from the in- stitution responsible for dataset creation, while represented country is assigned only when the dataset documentation identifies a target popula- tion or geographic context. For example, a model- generated Swahili dataset explicitly designed to simulate Kenyan speakers can be attributed to Kenya, whereas synthetic language-model output with no stated target population remains geographi- cally unattributed. A.4 Language Normalization and Inference Language metadata is often noisy, inconsistent, or incomplete. AtlasNLP therefore normalizes lan- guage metadata before using it for analysis or ge- ographic inference. The normalization process in- cludes whitespace and formatting cleanup, alias resolution, ISO-code expansion, preservation of locale-specific labels, separation of canonical lan- guages from language varieties and sign languages, and removal of invalid or non-language entries. Programming-language labels and other artifacts are excluded from the audited language inventory. Language normalization is conceptually separate from country attribution. A normalized language label does not automatically imply a represented country. Instead, language may contribute to the inferred geographic layer only when it provides a sufficiently specific geographic signal. Explicit country attribution is never assigned from language alone. Table 6 summarizes the final conservative language-to-country policy. The final attribution layers preserve this uncertainty explicitly.Among the 13,462 AtlasNLP-Core dataset records, 2,447 have explicit represented-country attribution, 1,059 have inferred-only attribution, and 9,956 remain unattributed. Thus, the broader explicit+inferred sensitivity layer contains 3,506 dataset records. In- ferred attribution is not synonymous with language- based mapping: it may also reflect other indirect Language signalTreatment Country-specific variety or locale May support inferred attribution when the language label itself pro- vides a sufficiently narrow geo- graphic signal, such as Egyptian Ara- bic for Egypt. Narrow geographic association May support conservative inferred attribution when the language or speech community is strongly asso- ciated with a limited geographic con- text. If multiple countries remain plausible, AtlasNLP does not force assignment to a single country. Language plus con- textual evidence May support inferred attribution when language is accompanied by additional country-specific evidence, such as a locally defined community, source, or collection context. Broadly distributed language No country is assigned from lan- guage alone.This includes lan- guages such as English, French, Spanish, Portuguese, Arabic, and Swahili; country-specific context is required. Script-only or am- biguous label No represented country is assigned when the available metadata identi- fies only a script, unresolved variant, or otherwise insufficient geographic signal. Table 6: Conservative treatment of language signals in represented-country inference. Language may sup- port the inferred layer when geographic evidence is sufficiently specific, but language alone never produces explicit country attribution. but geographically informative evidence described in Appendix A.3. The audited-language outputs are retained inde- pendently of these geographic decisions. The lan- guage audit was performed on the extracted records and then restricted to the final eligible Core and reviewed Gold collections, preserving the origi- nal language-level audit decisions after the post- extraction dataset audit. Final language inventory statistics are reported in Appendix B.2. A.5Automated Extraction for AtlasNLP-Core AtlasNLP-Core is constructed from ACL Anthol- ogy publications through a multi-stage candidate- selection and metadata-extraction pipeline. We begin with 119,963 ACL Anthology bibliography records spanning 1952â2025. Of these, 71,847 con- tain non-empty abstracts and are eligible for the dataset-paper filtering stage. We use ModernBERT-base-NLI to evaluate whether each paperâs title and abstract support the hypothesis: âThis paper proposes a dataset.â The NLI model assigns 23,674 papers an entail- ment label. We retain the 20,277 papers whose en- tailment score also exceeds 0.5 as candidate dataset papers for downstream extraction. Candidate papers are then processed using a staged extraction pipeline. For each paper, the pipeline retrieves the paper PDF, extracts full text page-by-page, performs an initial rule-based dataset-role triage, and applies schema-constrained GPT-4o-mini extraction to recover dataset meta- data. The extraction stage records dataset identity, dataset relationship, task category, language, licens- ing, represented-country signals, modality, syn- thetic status, format, and related metadata. Auto- matically extracted fields are accompanied by page- and quote-level evidence when available. Records requiring additional clarification are routed to a targeted second LLM pass using relevant paper excerpts rather than being accepted directly. The initial role-triage step excludes 905 can- didate papers before LLM extraction. A further 1,337 candidates fail during PDF retrieval, parsing, schema validation, or related extraction steps. The remaining 18,035 papers yield successful initial paper-level extractions: 20,277 = 905 + 1,337 + 18,035. These 18,035 records form the initial extrac- tion set, rather than the final AtlasNLP-Core col- lection. As described in Appendix A.7, a subse- quent 400-record human audit identified remain- ing dataset-role and represented-country attribution errors. The resulting audit criteria were then ap- plied across the full initial extraction set, yielding the final 13,462 AtlasNLP-Core primary dataset- contribution records. A.6 Metadata Extraction Prompt The LLM-assisted extraction stage uses a schema- constrained prompt designed to minimize unsup- ported inference and preserve auditable provenance. The model is instructed to return valid JSON, mark unavailable fields asNot stated, and pro- vide page-level evidence for extracted metadata. This prompt governs the initial extraction stage; dataset-role eligibility and represented-country at- tribution are subsequently re-evaluated using the post-extraction audit procedure described in Ap- pendix A.7. A simplified version of the extraction prompt is shown below. System prompt: You extract dataset metadata from the provided paper text conservatively and auditably. Rules: - Do not guess. Do not infer beyond the available evidence. - If a field is not supported by the paper text, output "Not stated". - Every non-"Not stated" field MUST have an evidence item with a page number and a short quote (<= 20 words). - Output MUST be valid JSON only. - If multiple datasets are described, identify the primary target dataset when possible; otherwise flag ambiguity. - All required fields must be present. User prompt: Return JSON with this structure: paper:..., dataset:..., evidence:[...], quality: flags:[], needs_stage2:boolean , run:... Allowed task categories: [ALLOWED_TASK_CATEGORIES] Paper ID: [PAPER_ID] Paper URL: [PAPER_URL] Prompt version: [PROMPT_VERSION] Paper text with page markers: [PAPER_TEXT_WITH_PAGE_MARKERS] A.7 Post-Extraction Audit and Refinement The 18,035 successful initial extractions were de- signed for high recall and still contained cases in which the paper did not make a primary dataset con- tribution or in which represented-country evidence was insufficiently grounded in dataset provenance. We therefore conducted a human audit of 400 ran- domly sampled records from the initial extraction set. The audit reviewed two issues separately. First, reviewers determined the relationship between the paper and the target dataset: whether the paper introduced a new dataset, materially extended an existing dataset, constructed a new dataset compi- lation, reused an existing resource, or did not sup- port a confident target-dataset identification. Sec- ond, reviewers evaluated represented-country at- tribution against evidence about the dataset itself, rather than author affiliation, language alone, or incidental country mentions. The audit identified recurring failure modes in both dimensions. In particular, some initial records described reused datasets rather than primary Final eligible relationshipDataset records Dataset introduced in paper12,205 Existing dataset materially extended819 New dataset compilation constructed438 Final AtlasNLP-Core13,462 Table 7: Composition of final AtlasNLP-Core by paperâ dataset relationship. Records are retained only when the target dataset is confidently matched and the paper makes a primary dataset contribution. dataset contributions, while some represented- country assignments reflected the location of au- thors or institutions rather than the population rep- resented by the dataset. Other errors arose from broadly distributed languages, incidental country mentions, incomplete multi-country scope, and ambiguous dataset identity. These findings were used to tighten both dataset-role and geographic- attribution criteria before re-evaluating the full 18,035-record initial extraction set. Final dataset eligibility. A record is retained in AtlasNLP-Core only when the target dataset can be matched with sufficient confidence and the paper makes one of three primary dataset contributions: (i) introducing the dataset, (i) materially extend- ing an existing dataset, or (i) constructing a new dataset compilation. Mere reuse of an existing dataset does not qualify. Table 7 summarizes the resulting final Core com- position. Of the 4,573 initial extractions not retained in fi- nal Core, 4,071 involve reuse of an existing dataset, 408 do not yield a sufficiently identifiable tar- get dataset, and 42 involve unresolved ambiguity among multiple datasets. An additional 48 records have a potentially eligible dataset relationship but lack a sufficiently confident match to the target dataset, and four records could not be processed successfully during re-evaluation. Thus, the re- duction from 18,035 initial extractions to 13,462 final Core dataset records reflects dataset-role and identity refinement rather than removal based on geographic coverage. Geographic refinement.Dataset eligibility is de- termined independently of whether represented- country metadata can be recovered.Eligible records remain in AtlasNLP-Core even when no country can be established. For geographic analyses, represented-country evidence is instead assigned to the separate explicit, inferred, or unattributed layers described in Appendix A.3. This separation prevents uncertainty in geographic provenance from being conflated with whether a paper contributes a dataset. The same 400-record audit is subsequently used to evaluate whether the revised pipeline addresses the failure modes observed during review. Post- revision audit results, including country-attribution corrections and abstention behavior, are reported in Appendix C. A.8 Producer-Country Extraction AtlasNLP-Core also records producer geography, defined as the institutional location associated with dataset production. Producer country is derived from author affiliations rather than researcher na- tionality, identity, or represented-country informa- tion. It is therefore treated as conceptually separate from represented-country attribution. The producer-side pipeline combines rule-based affiliation extraction with targeted LLM fallback. For each ACL paper, the pipeline retrieves affilia- tion information from the paper, including first- page text where available, and uses determinis- tic methods to identify institution names, explicit country mentions, email domains, and other affilia- tion cues. Country assignment follows a hierarchy of evidence that includes direct country mentions, institutionâcountry mappings, and domain-based signals. Extracted country names are normalized to the common 197-entity AtlasNLP ontology. When rule-based extraction leaves producer metadata missing or ambiguous, a targeted LLM fallback is applied. Trigger conditions include missing country or institution information and un- resolved or conflicting affiliation evidence. Rule- based results are retained when sufficiently sup- ported, while the fallback is used to resolve or fill remaining gaps. In the final AtlasNLP-Core collection, producer- country metadata is available for 13,224 of 13,462 dataset records (98.2%). We independently eval- uate this pipeline on 100 human-labeled records. At the country-association level, producer-country extraction achieves 96.9% precision and 97.7% re- call, and the primary producer country is recovered in all 100 audited records. Full audit results are reported in Appendix C. A.9 CountryâTask Matrix Construction Country-task matrices are constructed from the fi- nal 13,462-record AtlasNLP-Core collection after country and task normalization. Each Core dataset record has one normalized task category. For geo- graphic analyses, records are expanded over all represented countries supported by the relevant attribution layer, so that a multi-country dataset contributes once to each supported country while remaining a single dataset record in the underlying Core collection. We construct two versions of the matrix. The pri- mary matrix uses only explicit represented-country attribution.This layer contains 2,447 dataset records, which expand to 4,421 country-record as- sociations covering 158 countries. The sensitivity matrix additionally includes inferred geography, yielding 3,506 dataset records and 6,001 country- record associations across 168 countries. For comparability across geographic layers, both matrices are represented over the complete AtlasNLP-Core analysis space of 197 geopoliti- cal entities and 30 task categories, including zero- count rows and columns where applicable. Each matrix cell therefore records the number of country- record associations for a given country and normal- ized task. The explicit matrix contains 1,231 nonzero country-task cells out of 5,910 possible cells, leav- ing 79.2% of the matrix empty. Under the broader explicit+inferred attribution layer, 1,480 cells are nonzero and 75.0% remain empty. Thus, the broader attribution layer increases observed geo- graphic coverage but does not substantially alter the overall sparsity of country-task representation. These matrices underpin the country-task cover- age and task-portfolio analyses reported in the main paper. Unless otherwise stated, primary results use the explicit matrix, while the explicit+inferred ma- trix is used as a sensitivity analysis. B Data Transparency This appendix summarizes the controlled geo- graphic and language vocabularies used in At- lasNLP, coverage under the different geographic- attribution layers, and the audited language inven- tory. B.1 Country Ontology and Coverage AtlasNLP uses a standardized ontology of 197 geopolitical entities: 193 UN member states, the two UN non-member observer states, and Taiwan and Kosovo. The observer states are represented in the ontology as Palestine and Vatican City. Country names are normalized to this ontology before anal- ysis, reducing variation from aliases, abbreviations, alternate spellings, and sub-national references. SubsetExplicitExplicit + inferred AtlasNLP-Core 158 / 197168 / 197 AtlasNLP-Gold 192 / 197197 / 197 Table 8: Geopolitical-entity coverage under the two represented-country attribution layers. Primary analyses use explicit attribution; explicit+inferred coverage is reported as a sensitivity analysis. Table 9 lists the complete ontology. Coverage depends on the geographic-attribution layer being considered.Table 8 therefore re- ports coverage separately for explicit attribu- tion and for the broader explicit+inferred sen- sitivity layer. AtlasNLP-Core provides explicit represented-country evidence for 158 of the 197 entities; including conservative inferred attribu- tion increases coverage to 168. AtlasNLP-Gold provides substantially broader geographic cover- age, with explicit evidence for 192 entities and explicit+inferred coverage for all 197. The difference between ontology membership and observed coverage is intentional. Countries without attributed dataset records remain part of the 197-entity analysis space and therefore appear as zero-count cases rather than being dropped from country-level analyses. B.2 Language Inventory and Audit AtlasNLP records both raw language metadata and audited language labels. Raw fields may contain ISO codes, aliases, dialect or locale labels, script names, formatting variants, and non-language arti- facts. We therefore normalize and audit language metadata before using it for language-level anal- ysis. The normalization and conservative use of language in geographic inference are described in Appendix A.4. The language audit was conducted on the source collections and then aligned to the final post-audit Core and Gold records. All final records were suc- cessfully matched to their audited-language anno- tations. Programming-language artifacts identified during the audit, including Ada and ABAP, are ex- cluded from the published language inventory. InventoryAudited languages AtlasNLP-Core1,239 AtlasNLP-Gold269 AtlasNLP overall1,245 The Core and Gold inventories overlap substan- tially but are not identical. Six audited languages occur in Gold but not in final Core: Basaa, Guinea- Bissau Creole, Hiri Motu, Inuinnaqtun, Myene, and Nzebi. The union of the two final inventories there- fore contains 1,245 audited languages. These language counts describe normalized lan- guage coverage rather than geographic coverage. In particular, the presence of a language label does not imply that a represented country can be estab- lished for the corresponding dataset record. B.3 Language-to-Country Mapping Transparency Language metadata is not treated as direct evidence of represented-country coverage. In particular, At- lasNLP does not convert normalized language la- bels into countries through a general languageâ country lookup. Many languages span multiple countries, and assigning every dataset in a lan- guage to all countries where that language is spo- ken would substantially overstate geographic rep- resentation. Language can contribute only to the inferred geographic layer when the available label or ac- companying context provides a sufficiently specific geographic signal. For example, a country-specific variety such as Egyptian Arabic may support in- ferred attribution to Egypt. By contrast, broadly dis- tributed languages such as English, French, Span- ish, Portuguese, Arabic, or Swahili do not support country assignment on their own. Script-only la- bels, unresolved varieties, and cases compatible with several countries likewise remain geograph- ically unattributed unless additional evidence is available. Language-derived evidence never produces ex- plicit attribution. Explicit represented-country as- signments require direct provenance connecting the dataset content, participants, sources, collection process, or target population to a country. The com- plete inference policy is described in Appendix A.4. This separation allows AtlasNLP to retain rich lan- guage metadata without treating language coverage as equivalent to geographic coverage. C Validation and Quality Checks We evaluate AtlasNLP-Core through three comple- mentary validation procedures. First, we compare final Core metadata against independently curated AtlasNLP-Gold records that can be aligned with high confidence. Second, we revisit the 400-record human audit used to motivate the post-extraction re- finement and measure whether the revised pipeline corrects the observed dataset-role and represented- country attribution errors. Third, we conduct a sep- arate 100-record human audit of producer-country extraction. Together, these checks evaluate dataset identity, task assignment, represented-country attri- bution, abstention behavior, and producer geogra- phy. C.1 Gold-Core Alignment AtlasNLP-Gold and AtlasNLP-Core are con- structed through different pipelines, and a shared paper does not necessarily imply a shared dataset record. Papers may introduce or discuss multiple datasets, and dataset names can vary across sources. We therefore perform validation only after estab- lishing a high-confidence dataset-level alignment rather than treating paper-level overlap alone as a match. We begin with the 1,480 post-audit Gold entries marked as complete. Among these, 339 unique pa- pers have ACL Anthology links and can therefore be compared directly with the ACL-derived Core pipeline. Of these papers, 299 are recovered in the initial 18,035-record extraction set, and 267 remain after the final dataset-role and identity refinement that produces the 13,462-record Core. The 32 re- covered papers that do not remain in final Core consist of 28 cases classified as reuse of an existing dataset and four for which the target dataset could not be identified with sufficient confidence. Alignment stageCount ACL-linked Gold papers339 Recovered in initial extraction set299 Retained in final Core267 High-confidence dataset matches225 Within the 267 papers retained in final Core, we then establish dataset-level correspondence using conservative identity criteria. We first match the normalized Gold and Core dataset names within the same ACL paper. This yields 203 high-confidence matches. When normalized names do not match, we accept a record only if Gold and Core provide the same non-empty dataset URL within the same paper, yielding 22 additional matches. We do not use paper-URL agreement alone to infer dataset identity. The resulting validation set therefore contains 225 high-confidence Gold-Core dataset matches: 203 based on exact normalized dataset-name agree- ment and 22 based on exact dataset-URL agree- ment. These matched records form the reference population for the represented-country and task comparisons reported in Appendix C.2. C.2 Represented-Country and Task Validation We evaluate represented-country agreement on the 225 high-confidence GoldâCore dataset matches described above. Because the final pipeline dis- tinguishes explicit from inferred geography, we perform the comparison separately under the two attribution layers used in the main analysis. For each layer, we distinguish between (i) the number of Gold reference records with geographic evidence and (i) the subset for which Core also produces a represented-country attribution. Agree- ment metrics are computed conditional on Core making a geographic prediction. A Core abstention is therefore treated as missing geographic coverage rather than as an incorrect country assignment. MetricExplicitExplicit + inferred Gold reference records181225 Core predicts geography87 (48.1%)138 (61.3%) Any country overlap97.7%97.8% Exact country-set match66.7%68.8% Country-assignment precision94.7%94.0% Under the explicit-only comparison, 181 matched Gold records have explicit represented- country evidence. Core assigns explicit geography to 87 of these records. Conditional on an assign- ment, the Core and Gold country sets overlap in 97.7% of cases, and 66.7% have an exact country- set match. At the individual country-assignment level, precision is 94.7%. Under the broader explicit+inferred comparison, all 225 matched Gold records have geographic ev- idence in the corresponding Gold layer. Core as- signs explicit or inferred geography to 138 records. Conditional on an assignment, 97.8% have at least one country in common with Gold and 68.8% have an exact country-set match. Country-assignment precision is 94.0%. Exact-set agreement is deliberately strict. For a multi-country dataset, the metric is counted as cor- rect only when the complete predicted country set matches the Gold country set exactly; both omitted and additional countries cause the record to fail exact-set agreement. The substantially higher over- lap scores therefore indicate that most non-exact cases reflect differences in multi-country scope rather than completely unrelated geographic attri- bution. Task agreement.Task labels are compared only when the Gold record has a single task label that maps directly to the final Core taxonomy. Of the 225 high-confidence dataset matches, 219 satisfy this criterion. Core assigns the same normalized task category as Gold in 175 cases, yielding 79.9% exact task agreement. These results support the use of the automated metadata for aggregate country- and task-level anal- ysis while also motivating the conservative treat- ment of missing geography and the explicit separa- tion of inferred attribution from the primary explicit layer. C.3 Post-Revision Audit Verification We next return to the 400 records used in the post- extraction human audit and evaluate their outcomes under the revised pipeline. The purpose of this check is different from the GoldâCore comparison above: rather than measuring agreement with an independent dataset collection, it tests whether the specific dataset-role and geographic-attribution fail- ure modes identified during the audit persist after full re-evaluation. The human audit classifies 287 of the 400 sam- pled records as eligible primary dataset contribu- tions, 112 as ineligible, and one as a processing failure. Human-audit outcomeRecords Eligible primary contribution287 Excluded112 Processing failure1 Total400 Country-attribution corrections. The original audit labels 187 represented-country assignments as incorrect. After the audit-driven re-evaluation, 177 of these cases (94.7%) no longer remain in the explicit geographic layer. Similarly, among 85 cases for which the human audit judged the geographic connection to be inferred rather than explicit, 83 (97.6%) are not incorrectly promoted to explicit attribution. A particularly important test concerns absten- tion. Among the 287 dataset-eligible audit records, 98 cases have no represented country supported by the human review. The revised pipeline leaves 97 of these 98 records geographically unattributed. This indicates that the revised procedure largely dis- tinguishes unsupported geography from extraction failure rather than forcing a country assignment whenever country-related information appears in a paper. Observed failure modes.The audit revealed sev- eral recurring reasons for unsupported country as- signments. For example, the initial extraction as- signed the United States to the Non-Cooperative Guess What?! Dataset, reflecting the creatorsâ Uni- versity of Pittsburgh affiliation; however, the paper identifies the participants only as native English speakers and does not establish their country. The revised record therefore remains geographically unattributed. Similarly, MADial-Bench was ini- tially associated with China, the United Kingdom, and Australia, but the dataset is synthetically con- structed without direct evidence that these coun- tries constitute represented populations; the revised pipeline abstains from country attribution. Even- tRelBench provides another example in which an incidental country signal led to an initial Japan assignment despite insufficient dataset-level prove- nance; this assignment is likewise removed. Theseexamplesillustratewhycountry- attribution errors cannot be corrected simply by improving country-name extraction.The relevant distinction is whether a geographic signal describes the datasetâs represented population or content rather than an author affiliation, conference location, incidental mention, language association, or other contextual information. The post-revision audit therefore supports the conservative design used in the final analyses: un- supported assignments are preferentially removed or downgraded from explicit attribution, while un- certainty is preserved through the separate inferred and unattributed layers. C.4 Producer-Country Validation We independently evaluate producer-country ex- traction on a manually reviewed sample of 100 AtlasNLP-Core records. Human reviewers inspect author affiliations and record the producer-country set associated with each paper. The automated out- put is then compared against this reference at both the country-association and record levels. Across the 100 audited records, the extractor pre- dicts 130 producer-country associations, of which 126 are correct. The human reference contains 129 associations, of which 126 are recovered. This yields 96.9% precision and 97.7% recall. At the record level, the complete producer- country set matches the human reference exactly for 93 of 100 papers. Importantly, the primary producer country is recovered in all 100 audited records, including cases where additional sec- ondary affiliations cause the full country set to differ. Producer-country audit metricResult Country-assignment precision96.9% Country-assignment recall97.7% Exact country-set match93 / 100 Primary producer country recovered100 / 100 The audit also allows us to inspect the two ex- traction paths separately. Among the 60 records resolved primarily through deterministic affiliation parsing, 51 have an exact producer-country-set match. All 40 records routed through the targeted LLM fallback have an exact country-set match in the audit. These results suggest that the fallback is useful for affiliation structures that remain unre- solved after rule-based parsing, while the combined pipeline provides high producer-country coverage without relying exclusively on LLM inference. Because producer geography is derived from author affiliations, this audit evaluates only the producer side of AtlasNLP. It does not validate represented-country attribution, which is evaluated separately in Appendices C.2 and C.3. D Human Annotation and Validation Process This appendix describes the curation and validation process used to construct AtlasNLP-Gold. Gold was designed to broaden geographic coverage be- yond the ACL-derived Core collection, establish the annotation schema, and provide an indepen- dently curated reference for validation. The collec- tion originated through collaborative human cura- tion and was subsequently extended and audited with machine assistance under the same curation criteria. D.1 Contributor Assignment Collaborators selected countries from a shared as- signment sheet. Countries were grouped into high- , medium-, and low-resource levels based on ex- pected NLP dataset availability. The resource levels were used to balance annotation workload rather than to characterize countries intrinsically. Resource level (credit)Expected workload High (5)Many datasets; higher deduplica- tion burden Medium (3)Moderate expected dataset avail- ability Low (1) Few or potentially no identifiable datasets Collaborators were asked to select enough coun- tries to reach a target workload and were encour- aged to choose countries across multiple regions and resource levels. When possible, contributors selected countries for which they had geographic, linguistic, or domain familiarity. This design aimed to distribute the curation workload while support- ing careful country-specific search. D.2 Dataset Search Procedure For each assigned country, collaborators searched for existing NLP datasets using academic and public dataset sources.Suggested sources in- cluded ACL Anthology, Hugging Face, Papers With Code, GitHub, Kaggle, Google Scholar, Se- mantic Scholar, national data portals, institutional repositories, community projects, and regional dataset collections. Contributors prioritized official dataset sources such as repository pages, bench- mark websites, Hugging Face pages, GitHub repos- itories, DOIs, and associated papers. A shared taskâcountry matrix was used to re- duce duplicate work. Before entering a dataset, contributors checked whether the resource had al- ready been recorded. Existing entries were cross- checked rather than entered independently as new resources; newly identified datasets were added to the contributorâs entry sheet and shared tracking matrix. For lower-resource settings, contributors could also report cases in which no relevant dataset was identified after search. These reports documented the sources searched and supported the broader goal of distinguishing an unsuccessful search from an unexamined country. They are not interpreted as evidence that no NLP resource exists for that country. D.3 Human Data Entry Human-curated entries followed a shared At- lasNLP schema. Contributors recorded dataset name, task category, language, represented country, dataset and paper links, modality, multiple-choice format, synthetic status, license, provenance notes, and geographic-attribution evidence. Gold was initially organized around datasetâ country annotation instances. Multi-country re- sources could therefore be associated with more than one country while retaining a common dataset identity. For final reporting, dataset names are ad- ditionally normalized into dataset-name groups; these normalized groups are used only as an ap- proximate identity summary rather than as a claim of complete entity-level deduplication. D.4 Task Annotation Contributors assigned datasets to task categories using a shared guide based on ACL and EMNLP thematic areas. The guide included broad task categories and example subtasks.For exam- ple, information extraction included named en- tity recognition, relation extraction, event extrac- tion, keyphrase extraction, document retrieval, and knowledge-base population.Question answer- ing included extractive, abstractive, open-domain, multi-hop, multiple-choice, and conversational QA. Multimodality included image captioning, visual question answering, videoâtext alignment, and re- lated language-grounded tasks. If a dataset did not fit an existing category, con- tributors could propose a new category. These addi- tions were subsequently reviewed and normalized to the final AtlasNLP taxonomy described in Ap- pendix A.2. D.5 Human Country Attribution Represented-country annotation followed the same conceptual distinction used in the final AtlasNLP analyses. Contributors prioritized direct evidence that the dataset content, participants, data sources, collection setting, or target population was asso- ciated with a country. Such evidence supports ex- plicit attribution. When direct country information was unavail- able, geographically informative but indirect evi- dence could support an inferred attribution. Lan- guage alone was treated conservatively. Broadly distributed languages such as English, French, Spanish, Portuguese, Arabic, and Swahili were not considered sufficient for country attribution with- out additional country-specific evidence. Cases lacking adequate explicit or conservative inferred evidence remained unattributed. For multi-country datasets, annotators retained all countries supported by the available evidence rather than forcing a single primary geographic label. Country-attribution decisions were recorded separately from producer geography. D.6 Cross-Validation After initial compilation, entries were redistributed for secondary validation. Validators performed tar- geted checks rather than re-annotating every source from scratch. Validation focused on four areas: ⢠Dataset and link checks: Validators con- firmed that dataset names corresponded to the linked resources and attempted to repair bro- ken or incorrect links when possible. ⢠Core metadata checks: Validators checked country, language, task, and geographic- attribution evidence for consistency with the cited resource. â˘Missing-value checks: Missing fields were completed when the relevant information could be established with high confidence. â˘Formatting and normalization checks: Validators corrected invalid controlled- vocabulary values, inconsistent country names, and formatting errors. Validators recorded comments and changes in provenance or notes fields and marked entries com- plete after validation. D.7 Country Name Normalization Country names were normalized to the 197-entity AtlasNLP ontology described in Appendix B.1. Common corrections included mapping âUAEâ to âUnited Arab Emirates,â âAmericaâ to âUnited States,â âCzech Republicâ to âCzechia,â âRepublic of Congoâ to âCongo (Congo-Brazzaville),â and âIvory Coastâ to âCote dâIvoire.â Where the ontology required country-level rep- resentation, sub-national labels were normalized to the corresponding geopolitical entity; for example, Wales and Scotland were mapped to the United Kingdom and Puerto Rico to the United States. For multi-country entries, normalization was applied to individual country labels while preserving the remaining country set. D.8 Machine-Assisted Extension and Final Audit Following the original collaborative curation, AtlasNLP-Gold was extended and re-audited with machine assistance. These stages followed the same underlying curation goals and geographic- attribution criteria as the original annotation pro- cess. Machine assistance was used to surface can- didate evidence, identify records requiring recon- sideration, and support consistency checks; it did not change the definition of represented-country evidence. The subsequent audit focused especially on records for which country attribution depended on indirect evidence, dataset identity was unclear, or the available provenance did not support the exist- ing geographic assignment. Unsupported assign- ments were removed, while geographically plau- sible but indirectly supported cases were retained separately as inferred attribution. This process also aligned Gold country names, attribution labels, and language metadata with the final controlled vocab- ularies used in Core. D.9 AtlasNLP-Gold Output The final AtlasNLP-Gold collection contains 1,480 curated entries corresponding to 989 normal- ized dataset-name groups. Gold provides explicit represented-country coverage for 192 of the 197 AtlasNLP entities; including conservative inferred attribution expands this to all 197 entities. Because Gold was deliberately curated to broaden geographic coverage, its country distri- bution should not be interpreted as a naturally oc- curring sample of the NLP dataset ecosystem. Its primary roles are complementary coverage, meta- data transparency, and independent reference data for validating AtlasNLP-Core. E Dataset Examples Table 10 provides representative examples of the geographic metadata cases distinguished in At- lasNLP: explicit multi-country coverage, separa- tion of represented and producer geography, con- servative abstention when represented-country evi- dence is unavailable, and inferred geography for a synthetic resource. F Hugging Face Comparison Hugging Face is the largest public repository for NLP datasets, with over 900,000 datasets, includ- ing approximately 450,000 associated with text modalities. However, many datasets lack the struc- tured metadata required for systematic analysis. Language and task annotations are missing in over 80-85% of records, and no fields capture repre- sented country or represented population. Al- though anuploader_regionfield is widely pop- ulated, its distribution suggests that it reflects account defaults rather than dataset provenance. Datasets linked to research papers tend to be more consistently documented, but their proportion has declined as platform growth has shifted toward in- dustry and individual contributors. These patterns reflect a trade-off between scale and structure. ACL-derived dataset records provide more consistent metadata but more limited cover- age, while Hugging Face offers much greater scale without standardized annotations for geographic representation. This lack of structured metadata highlights the need for resources such as AtlasNLP, which organize existing dataset contributions using a standardized metadata framework and can inform the design of future dataset platforms. G Robustness and Sensitivity Analyses Primary geographic analyses in the paper use only explicitly supported represented-country attribu- tion. To assess whether the main conclusions de- pend on this conservative choice, we repeat key analyses using the broader explicit+inferred at- tribution layer. This section also reports addi- tional robustness checks addressing multi-country resources, dataset identity, and pipeline recovery. G.1 Sensitivity to Geographic-Attribution Layer Including conservative inferred geography in- creases the number of Core dataset records with represented-country metadata from 2,447 to 3,506 and expands observed country coverage from 158 to 168 of the 197 AtlasNLP entities. Despite this increase, the principal geographic patterns remain similar. StatisticExplicit Explicit + inferred Dataset records with geography2,4473,506 Countries represented158168 Empty countryâtask cells79.2%75.0% Infrastructure Pearson r0.650.70 Infrastructure Spearman Ď0.610.67 Countries in task-portfolio analysis7789 Median task breadth1113 Median top-3 task share55.6%54.3% Country-task coverage remains highly sparse: 79.2% of cells are empty under explicit attribution and 75.0% remain empty after inferred geography is included. The association between represented- record availability and research infrastructure is also similar and slightly stronger under the broader layer, with Pearson correlation increasing from ap- proximately 0.65 to 0.70 and Spearman correlation from 0.61 to 0.67. Task-portfolio patterns are likewise stable. Among countries with at least 10 represented dataset records, the number of eligible countries increases from 77 to 89, while median task breadth increases from 11 to 13 of the 30 Core task cat- egories. Median concentration in the three most common tasks changes little, from 55.6% to 54.3%. Production-representation asymmetry also per- sists. Without a minimum-count restriction, 60 of 89 countries (67.4%) under explicit attribution and 65 of 96 countries (67.7%) under explicit+inferred attribution have content self-representation below 0.5. The primary results use the more conservative minimum-count analysis, under which 39 of 62 countries with at least 10 represented records with producer metadata and 10 producerârepresentation associations are primarily represented through records produced by institutions outside the coun- try. Altogether, these comparisons indicate that adding conservative inferred geography increases coverage but does not materially change the main conclusions regarding country-task sparsity, production-representation asymmetry, research- infrastructure alignment, or task concentration. G.2 Additional Robustness Checks Broad multi-country records.Because records spanning many countries can contribute dispro- portionately to expanded country counts, we re- peat the represented-country analysis after exclud- ing records attributed to more than 20 countries. Under explicit attribution, only 10 of 2,447 at- tributed records (0.4%) exceed this threshold, ac- counting for 234 of 4,421 country-record associ- ations (5.3%). After removing them, 4,187 asso- ciations remain across 154 countries. The top-10 represented-country set is unchanged, and coun- try rankings remain highly stable (SpearmanĎ = .991). The top-10 share increases from 48.6% to 50.4%, indicating that very broad multi-country records slightly dilute, rather than produce, the ob- served concentration. Exact-name sensitivity. AtlasNLP-Core uses paper-level dataset-contribution records rather than globally resolved dataset entities because aliases, versions, extensions, and generic dataset names make full entity resolution ambiguous. As a lim- ited sensitivity check, we collapse records sharing the same exact normalized dataset-name key. Of 13,462 Core records, 13,451 have a usable normal- ized key, corresponding to 13,131 distinct name groups. Within the explicit geographic layer, this collapse reduces 4,421 country-record associations to 4,399 name-groupâcountry associations across 2,427 geographically attributed name groups. The top-10 represented-country set remains identical; across the baseline top 50 countries, rankings cor- relate atĎ = .999, with a maximum rank shift of two positions. The top-10 share changes by only 0.05 percentage points (48.59% to 48.53%). Thus, repeated exact normalized-name matches do not materially drive the headline geographic concentra- tion results. This check is intentionally narrower than full dataset-entity resolution. 0.000.250.500.751.00 Content self-representation (share of dataset records about country X produced domestically) 0.0 0.2 0.4 0.6 0.8 1.0 Producer self-representation (share of producer representation associations involving a country that represent that same country) Producer vs. content self-representation explicit + inferred Regime Locally produced, self-focused Locally produced, outward-facing Externally represented, self-focused Externally represented, outward-facing Hollow = producer_count < 10 Brazil China Denmark Egypt India Nigeria Russia United States Figure 7: Dataset production and representation by country under explicit+inferred attribution. The x-axis measures the share of datasets about a country produced domestically; the y-axis measures the share of producer- representation associations involving a country that represent itself. Dashed lines define four production- representation regimes. Point size reflects dataset vol- ume; hollow points indicate countries with fewer than 10 producer associations. Analytical units and low-count handling. Ta- ble 11 summarizes the principal analytical units used in the paper. Country-level coverage counts expanded country-record associations, so a record explicitly representingNcountries contributes one association to each. Production-representation anal- yses additionally expand records over producer countries. Country-task sparsity instead uses the fixed 197Ă 30 country-task space. Low Lower-mid Upper-mid High 10 0 10 1 10 2 10 3 Rep. dataset records (log scale) (A) Coverage by income group | explicit + inferred 10 1 10 0 10 1 10 2 10 3 # universities (log scale) 10 1 10 0 10 1 10 2 10 3 Rep. dataset records (log scale) United States India Nigeria (B) Coverage vs. university count | explicit + inferred Income Low Lower-mid Upper-mid High Figure 8: Relationship between represented-country dataset records and research infrastructure under ex- plicit+inferred attribution. (A) Coverage by World Bank income group. (B) Coverage increases with national university counts on log-transformed values. Pearson râ .70and SpearmanĎâ .67. Color indicates income group. Best viewed in color. Analysis setUnitN AtlasNLP-Coredataset records13,462 Explicit rep.records / country-record assoc.2,447 / 4,421 Explicit+inferred rep.records / country-record assoc.3,506 / 6,001 Production-rep. (explicit)records w/producer / expanded assoc.2,396 / 6,533 Country-task matrixcountry-task cells5,910 (197Ă 30) Table 11: Denominator map for the principal AtlasNLP- Core geographic analyses. Different analyses expand paper-level dataset records according to the geographic relation being measured. For production-representation comparisons, the primary summary restricts interpretation to coun- tries with at least 10 represented records with pro- ducer metadata and 10 producer-representation as- sociations. Lower-volume countries remain visible in the figure but are treated descriptively because their ratios can change substantially with only a few records. Section G.1 additionally reports re- sults without this minimum-count restriction and under explicit+inferred attribution. G.3 Recovery Diagnostic Because AtlasNLP-Core is constructed through ACL metadata retrieval, NLI screening, automated extraction, and a subsequent dataset-role audit, dif- ferences in pipeline recovery can affect analyses 0.00.20.40.60.81.0 Share of dataset records within language Chinese Portuguese French English Spanish Language Country concentration within language-labeled dataset records explicit + inferred China Taiwan Japan Brazil Portugal Mozambique France Switzerland Canada United States India United Kingdom Spain Mexico Peru Others Figure 9: Country concentration within language- labeled dataset records under explicit+inferred attribu- tion. Country representation varies substantially within widely used languages: some are dominated by one or a few countries, while others are more geographically dis- tributed. Language coverage therefore does not directly indicate which countries are represented. over publication time. We therefore treat tempo- ral patterns as descriptive and use AtlasNLP-Gold only as a reference-set diagnostic rather than as an estimate of population-wide recall. Among 339 unique ACL-linked Gold papers in the final curated collection, 299 (88.2%) were recovered in the initial 18,035-record extraction set, and 267 (78.8%) are retained in the final 13,462-record Core. Conditional on initial recov- ery, 267/299 (89.3%) survive the final dataset-role and provenance audit. Of the 32 initially recovered Gold papers subsequently excluded, 28 were clas- sified as reuse of an existing dataset and four as cases in which the target dataset contribution could not be established. StageGold papers recoveredRate ACL-linked Gold reference339â Initial extraction29988.2% Final AtlasNLP-Core26778.8% Final, conditional on initial recovery267 / 29989.3% Table 12: Reference-set recovery for ACL-linked AtlasNLP-Gold papers across the extraction and final au- dit stages. Rates are diagnostic of this curated reference set and should not be interpreted as population-wide ACL recall. Diagnostic checks conducted during pipeline de- velopment also showed that recovery varies across publication periods rather than following a uniform or monotonic pattern. Because both document availability and extraction behavior can vary with publication year, we do not interpret differences in the temporal composition of AtlasNLP-Core as evidence of changes in the underlying NLP dataset ecosystem. Accordingly, the camera-ready analysis Figure 10: Heatmap of dataset coverage across countries and tasks under explicit+inferred attribution. Coverage is highly fragmented, with most country-task pairs lacking represented dataset records and even high-coverage countries concentrated in a subset of tasks. Only a subset of country labels is shown for readability, with labels placed at regular intervals (approximately every fifth country). Best viewed in color. does not make substantive claims about temporal diversification of country task portfolios. Country/entityCountry/entityCountry/entityCountry/entity AfghanistanEcuadorLuxembourgSĂŁo TomĂŠ and PrĂncipe AlbaniaEgyptMadagascarSaudi Arabia AlgeriaEl SalvadorMalawiSenegal AndorraEquatorial GuineaMalaysiaSerbia AngolaEritreaMaldivesSeychelles Antigua and BarbudaEstoniaMaliSierra Leone ArgentinaEswatiniMaltaSingapore ArmeniaEthiopiaMarshall IslandsSlovakia AustraliaFijiMauritaniaSlovenia AustriaFinlandMauritiusSolomon Islands AzerbaijanFranceMexicoSomalia BahamasGabonMicronesiaSouth Africa BahrainGambiaMoldovaSouth Korea BangladeshGeorgiaMonacoSouth Sudan BarbadosGermanyMongoliaSpain BelarusGhanaMontenegroSri Lanka BelgiumGreeceMoroccoSudan BelizeGrenadaMozambiqueSuriname BeninGuatemalaMyanmarSweden BhutanGuineaNamibiaSwitzerland BoliviaGuinea-BissauNauruSyria Bosnia and HerzegovinaGuyanaNepalTaiwan BotswanaHaitiNetherlandsTajikistan BrazilHondurasNew ZealandTanzania BruneiHungaryNicaraguaThailand BulgariaIcelandNigerTimor-Leste Burkina FasoIndiaNigeriaTogo BurundiIndonesiaNorth KoreaTonga Cabo VerdeIranNorth MacedoniaTrinidad and Tobago CambodiaIraqNorwayTunisia CameroonIrelandOmanTurkey CanadaIsraelPakistanTurkmenistan Central African RepublicItalyPalauTuvalu ChadJamaicaPalestineUganda ChileJapanPanamaUkraine ChinaJordanPapua New GuineaUnited Arab Emirates ColombiaKazakhstanParaguayUnited Kingdom ComorosKenyaPeruUnited States Congo (Congo-Brazzaville)KiribatiPhilippinesUruguay Costa RicaKosovoPolandUzbekistan Cote dâIvoireKuwaitPortugalVanuatu CroatiaKyrgyzstanQatarVatican City CubaLaosRomaniaVenezuela CyprusLatviaRussiaVietnam CzechiaLebanonRwandaYemen Democratic Republic of the CongoLesothoSaint Kitts and NevisZambia DenmarkLiberiaSaint LuciaZimbabwe DjiboutiLibyaSaint Vincent and the Grenadines DominicaLiechtensteinSamoa Dominican RepublicLithuaniaSan Marino Table 9: AtlasNLP country ontology used for normalization and coverage analysis. The ontology contains 193 UN member states, the two UN non-member observer states, Taiwan, and Kosovo. Metadata fieldSALTAfriSpeech-DialogToxASCIICode-170k-lingala SubsetAtlasNLP-GoldAtlasNLP-CoreAtlasNLP-CoreAtlasNLP-Gold Task categoryMachine TranslationDialogue and Interactive SystemsEthics, Bias, and FairnessCode Models Language(s) English; Swahili; Luganda; Run- yankole; Acholi; Lugbara; Ateso EnglishEnglishLingala Represented countryUganda; Kenya; TanzaniaNigeria; Kenya; South AfricaUnattributedCongo(Congo-Brazzaville); Democratic Republic of the Congo; Angola; Central African Republic Producer countryNot consistently available in Gold United StatesFranceNot consistently available in Gold Country attributionExplicitExplicitUnattributedInferred SyntheticNoNoNoYes IllustratesExplicit multi-country regional resource Representedâproducergeo- graphic distinction Producer metadata without sup- ported represented geography Synthetic resource with inferred geographic relevance Table 10: Representative AtlasNLP records illustrating common geographic metadata cases. Gold sources extend beyond ACL Anthology and therefore do not consistently contain producer-country metadata. Inferred geography is kept separate from the explicit attribution used in primary analyses.