Glottography: an open-source geolinguistic data platform for mapping the world’s languages

Repository location

All Glottography datasets are openly available through the Zenodo Glottography community at https://zenodo.org/communities/glottography (Accessed: 2026-02-19). Individual DOIs for each dataset are provided in the DOI column of Table 1. The datasets are maintained via the Glottography organisation on GitHub,¹ where users can track changes over time, report issues, suggest improvements, and contribute updates. The Rglottography package (Ranacher, 2026b) provides a convenient interface for downloading Glottography datasets and importing them directly into the R programming environment. The package is developed and maintained on GitHub through the Rglottography² repository and archived through Zenodo to ensure long-term accessibility and versioned releases.

Repository name

The Glottography community on Zenodo (for official releases) and the Glottography organisation on GitHub (for maintenance and feedback).

Object name

All dataset names are listed in the Dataset name column of Table 1.

Format names and versions

Each dataset is provided in Cross-Linguistic Data Format (CLDF), a standard for historical and typological language data (Forkel et al., 2018). Each source is assigned its own repository, structured as follows:

• – The etc folder contains CSV files with attribute data and BibTeX files referencing the source publications.

• – The raw folder contains the speaker area polygons in GeoJSON format.

• – The cldf folder stores the CLDF datasets, which aggregate the speaker area polygons according to the classification in the source publication, as well as at the Glottolog language and language family levels.

Dataset creators

All datasets were created by the Glottography consortium, whose members are the authors of this publication. In addition, the Source publication column of Table 1 lists the authors of the original sources who provided the primary maps and data.

Language

English.

License

The datasets are published under a CC-BY-4.0 license.

Table 1 lists all currently available Glottography datasets together with their source publication, linguistic and geographic coverage, time period covered, number of languages and dataset name and DOI. The (spatial) overlap column indicates whether two or more language areas in the source publication can overlap geographically and share the same space on the map, with the value partly indicating that some, but not all, maps in the source contain overlapping polygons.

3.1 Georeferencing source maps

Georeferencing⁴ assigns geographic coordinates to either a language map image (formats i and ii) or vector geometries without spatial reference (format iii), enabling accurate alignment in a GIS. It was performed using the Georeferencer plugin in QGIS, where the unreferenced map (image or vector) is displayed alongside a reference basemap with landforms or administrative boundaries. Shared, recognisable features such as coastal bends or river estuaries were used to place control points across both layers. These control points aligned the two layers by shifting and warping the language map (language geometries) from its original position to its correct spatial location. If noticeable distortions appeared in specific regions, additional control points were added locally. If the CRS of the source map was known, we used it as the target CRS for the georeferenced map; otherwise, we applied a generic global CRS, e.g., Web Mercator (EPSG:3857) or WGS84 (EPSG:4326).

3.2 Digitising language polygons

Digitising⁵ involves tracing the outlines of language areas on the georeferenced map image and converting them into polygon geometries, and was needed for formats i and ii. The georeferenced language maps were exported as GeoTIFF and digitised using the Advanced Digitising toolbar in QGIS, following one of two approaches. Polygons were either traced from scratch using the Add Feature tool or derived by cutting them from existing geometries of the Earth’s landmasses and major islands (Natural Earth, 2024) using the Split Features tool. Adding features is generally faster and involves a simpler workflow in QGIS, whereas cutting from existing geometries ensures that language areas align with landmasses, coastlines, and major lake shores. Adding features was preferred for regional inland language maps (e.g., for hochstetler2004sociolinguistic), while splitting features was better suited to maps covering continental-scale or coastal language areas (e.g., for asher2007world). In both cases, we digitised the polygons in the CRS of the georeferenced map and later reprojected them to a common CRS (WGS84, EPSG:4326).

For format iv), where spatial polygons were already available, any apparent geometric issues — such as invalid geometries that failed to close — were corrected, but the polygons were otherwise left unchanged. Finally, all polygons from a single source representing the same language were aggregated. Aggregation resulted in a single polygon geometry if the polygons formed one contiguous area, or a MultiPolygon geometry — a combined geometry consisting of multiple disjoint polygons — if they were spatially separated. All (Multi)Polygons from a single source were saved in a single GeoJSON file.

3.3 Attributes and metadata

Glottography records relevant attributes and metadata⁶ to uniquely identify and describe each source publication, language map, and speaker area polygon depicted in each map. A citation key in the format authorYYYYtopic identifies the source publication. The citation key points to a scientific reference stored in BibTeX format, which includes bibliographic metadata, such as the author(s), year, type of publication, and title. References were retrieved from official publisher websites, Glottolog, or Google Scholar, with manual corrections made to address any missing or inaccurate entries.

For each digitised speaker area polygon, we recorded the following attributes: the language name as it appears on the map; a unique numeric id to uniquely identify the polygon; the year the language area refers to, which can be the date indicated on the source map or, if not explicitly given, the publication year of the source; the full map name(s) used to uniquely identify the language map(s) in the source publication; if a language area spans multiple maps, all map names are listed, separated by a vertical bar; the Glottocode, a unique identifier assigned to the language (see next section); and a note, for any additional comments or annotations about the language area, for example if the language area was poorly visible on the map.

3.4 Linking to Glottolog

Glottography uses Glottocodes⁷ to uniquely identify each language polygon. Glottocodes are standardised, unambiguous identifiers for language varieties, maintained by Glottolog (Hammarström et al., 2025). They provide a consistent way to reference languages, dialects, and language families. Although some publications, such as Haynie & Gavin (2019), already included Glottocodes, most other publications lacked them. For these, we assigned Glottocodes to the language areas following the approach in Ranacher et al. (2025). We used the Python 3 guess_glottocode package (Ranacher, 2025), which filters candidate languages based on spatial proximity and then identifies the most suitable candidate using large language models (LLMs) and web crawling, or, if necessary, manual annotation.

3.5 Data curation

Data curation⁸ aggregates the raw speaker area polygons according to the classification in the source publication, or by language or language family as defined in Glottolog, and exports them in CLDF format with all polygons in the WGS84 CRS (EPSG:4326). We used the pyglottography Python 3 package to create three sets of vector geometries in GeoJSON format, each enriched with Glottocodes at different levels of aggregation:

• – Features: Speaker areas retaining the classification from the source publication.

• – Language areas: Speaker areas aggregated at the language level according to Glottolog’s classification.

• – Family areas: Speaker areas aggregated at the top-level language families according to Glottolog’s classification.

3.6 Error correction

We addressed potential errors arising during data collection and performed additional quality checks, both automated and manual.⁹ For geometry validation, we used shapely’s is_valid function in Python 3. Name checks compared the language names on the map with those in Glottolog, including any known alternative names. Glottocode checks ensured that each code was present in the Glottolog database and conformed to the correct format. While we corrected errors introduced during data processing, we generally did not alter the original data except for clear mistakes. Disputable mappings were preserved, as each source was treated as a valid, subjective interpretation of a language’s geographic area, consistent with our data collection policy.

In a project on the scale of Glottography, error correction will likely be an ongoing endeavour. To address this, we implemented a workflow designed to make error identification and resolution straightforward.

4.1 Coverage

We assessed the geographic density of language polygons in the Glottography data using an equal-area global hexagonal grid (CRS: EPSG:8857; Šavrič et al. 2019) (Figure 1). For each grid cell, we counted the number of language polygons that intersected it, considering only unique languages. Density is generally highest in regions known for high language diversity, such as Papua New Guinea, West Africa, and the upper Amazon in Bolivia and Peru. Glottography incorporates languages from multiple sources and time periods, including extinct languages; therefore, polygon density should not be interpreted as a direct measure of actual language diversity.

Figure 1

We also evaluated the coverage across the 25 major language families by number of languages (Figure 2). Most families are well represented, but there are several notable exceptions with low coverage, including Indo-European, Dravidian, Otomanguean, Tai-Kadai, Pidgin, and Hmong-Mien. This highlights immediate future directions for extending coverage and filling data gaps.

Figure 2

4.2 Comparison with Ethnologue and Glottolog

We compared Glottography’s coverage with two established reference datasets: the paywalled polygons provided by Ethnologue (Eberhard et al., 2024) and the language point coordinates provided by Glottolog (Hammarström et al., 2025). Ethnologue includes a total of 7,651 unique polygons, compared to 5,338 language polygons in Glottography. To evaluate regional differences between the two datasets, we intersected the hexagonal grid with polygons from both sources, counted the number of unique languages in each grid cell, and computed the coverage difference (Glottography – Ethnologue) (Figure 3). Because Ethnologue uses the ISO 639-3 standard for language identification rather than Glottocodes, there is no fully compatible way to filter language polygons across both datasets. We compared the language polygons in Glottography with all entries in Ethnologue. Since some Ethnologue polygons are not classified as languages by Glottolog, this comparison tends to favour Ethnologue. For example, Serbian, Croatian, and Bosnian are treated as dialects in Glottolog and are therefore aggregated under the single language Serbian-Croatian-Bosnian in Glottography, whereas Ethnologue treats them as three distinct languages.

Figure 3

Overall, Glottography’s coverage closely matches that of Ethnologue. It even exceeds Ethnologue across large parts of the upper Amazon in Peru and Bolivia, the northwestern United States and Canada, and northern Australia. In southern Africa and northern Asia, Glottography generally matches, and in some areas slightly exceeds, Ethnologue’s coverage. The weakest regions are Mesoamerica, Papua New Guinea, West Africa, India, and Southeast Asia, where Glottography currently falls short of Ethnologue.

Because Glottolog provides only point coordinates rather than polygons, coverage per grid cell cannot be directly compared between the two sources. Instead, we indicate which of the languages with point coordinates in Glottolog have a corresponding polygon in Glottography (Figure 4). The results are consistent with those obtained when comparing Glottography to Ethnologue. Overall, Glottography includes polygons for 65% of the 8,300 languages with point coordinates in Glottolog. Notable gaps remain in Mesoamerica, Europe, Papua New Guinea, West Africa, India, and Southeast Asia, where Glottography lacks Glottolog languages.

Figure 4

4.3 Exemplary languages in Glottography

We argue that language maps should reflect the plurality of perspectives regarding a language’s location. To address this, Glottography incorporates multiple polygons per language, each sourced from different references.

We illustrate examples of language polygons from multiple sources for four languages (Figure 5): Shona (sub-Saharan Africa), Bengali (South Asia), Algonquin (North America), and Bulgarian (Europe). For Shona, asher2007world includes most of Zimbabwe, except for the southwest, which it assigns to Ndebele, and the areas of Zambia surrounding Lake Cahora Bassa. carling2025diacl largely follows asher2007world but additionally includes parts of Mozambique within the Shona area. The Shona polygon in wikipedia2024officiallang corresponds to the national territory of Zimbabwe.

Figure 5

There are four sources for Algonquin: haynie2019modern, goddard1999native, asher2007world, and carling2025diacl. They largely agree that the language was traditionally spoken northwest of present-day Montreal, within the provinces of Ontario and Quebec. goddard1999native differs in that it also includes a large area north of Montreal that is absent from the other sources.

For Bengali, asher2007world includes most of Bangladesh, as well as parts of India: to the west (West Bengal) and to the east (Tripura, Meghalaya, and Assam). carling2025diacl covers Bangladesh, West Bengal, and Tripura in India. bouckaert2012indoeuropean includes most of Bangladesh except for parts of Rangpur, Rajshahi, and Mymensingh in the north. grierson1903lsi covers Bangladesh, Tripura, Meghalaya, and Assam in India, and even the border area with Myanmar south of Chattogram. The Bengali polygon in wikipedia2024officiallang corresponds to the national territory of Bangladesh.

The Bulgarian language area in asher2007world includes parts of Bessarabia along the border between Moldova and Ukraine, but excludes several regions in central, southern, and northeastern Bulgaria, which are instead assigned to Gagauz. bouckaert2012indoeuropean similarly excludes these areas from Bulgarian, additionally assigns western Bulgaria to Macedonian, but omits Bessarabia. carling2025diacl largely follows the polygons in asher2007world, omitting the enclaves and likewise excluding Bessarabia. Finally, the polygon for Bulgarian in wikipedia2024officiallang corresponds to the national territory of Bulgaria.

At present, most languages in Glottography are represented by only one or two sources, while languages with more sources are relatively rare. As several publications focus on specific languages within a region—for example, exclusively Indo-European languages in Northwestern Europe—they contribute additional sources for some languages.

4.4 Aggregation and its implications

Glottography provides speaker areas at three levels of aggregation: as features, languages, and top-level families, as defined by Glottolog. Here, we briefly discuss aggregation and some of its implications.

The features always retain the classification of the original publication. For example, asher2007world views Danu and Intha in Myanmar as separate languages; hence, the features include separate speaker area polygons for both. In contrast, Glottolog views Danu (danu1251) and Intha (inth1239) as dialects of the Danu-Intha language (inth1238). When aggregating at the language level, the polygons for Danu and Intha are combined into a single Danu-Intha language area, following Glottlog’s classification.

Wikipedia lists Mongolian (mong1331) as the official language of Mongolia, whereas Glottolog treats it as a subfamily of Mongolic-Khitan, with daughter languages Halh Mongolian, Oirad-Kalmyk-Darkhat, and Peripheral Mongolian. When aggregating wikipedia2024officiallang at the language level, Mongolian is therefore excluded, since mapping it unambiguously to a single daughter language is not possible. More generally, polygons corresponding to Glottolog (sub-)families are omitted at the language level; this applies, for example, to Uzbek and Azerbaijani in wikipedia2024officiallang and to Malagasy in asher2007world. A special case is the traditional speaker areas of Australia in asher2007world, where multiple languages can share one area. We mapped these to the closest common subgroup in Glottolog at the feature level (see also Ranacher et al. 2025).

The choice of whether to use feature-, language-, or family-level speaker areas depends on the specific use case. We recommend using feature polygons when the classification in the source is deemed appropriate, or when the classification itself is not crucial to the task at hand. For example, when creating a map for Danu, it may not matter that asher2007world treats it as a language, whereas Glottolog classifies it as a dialect. When consistent classification is important, we instead recommend using language- and family-level polygons, which follow Glottolog’s classification of languages and (top-level) families. In all cases, we advise readers to consult the relevant Glottolog entry and inspect the corresponding polygon geometry to ensure that speaker areas at a given level of aggregation are appropriate for their use case.

5.1 Limitations and future work

We aim to establish Glottography as a free, open, and community-maintained repository for collecting digital areal information about languages, analogous to the role that Glottolog plays in language classification. To achieve this goal, current limitations must be addressed.

While more than 60% of languages in Glottolog have a polygon (Figure 4), coverage remains low in certain regions. We aim to expand the dataset by collecting additional language areas, particularly for geographic regions and time periods with limited coverage. We also seek to engage other researchers in refining the standards for Glottography and integrating their language areas of interest into the repository. Ideally, contributions will extend beyond the academic community, allowing also speakers of (minority) languages to map the areas of their own languages. A first step in this direction is the set of tutorials,¹⁰ which provide detailed instructions for collecting and contributing data to Glottography and are available to anyone wishing to participate in the project.

Glottography treats all scientific sources that meet the formal requirements for inclusion equally, with a strong emphasis on transparency: no source is considered inherently better than another. Currently, the most comprehensive dataset with full global coverage is asher2007world. Still, this dataset has limitations. Coverage is relatively sparse in some regions, notably Western Africa. In addition, the classification of linguistic varieties into separate languages, dialects, or subfamilies does not always align with that in Glottolog, for example in the case of Australian Aboriginal languages. For these, other datasets (e.g., hochstetler2004sociolinguistic, bowern2021australia) likely provide more suitable alternatives. A logical next step would be to compile a consensus dataset that selects the most appropriate language polygon for each language entry in Glottolog. However, given that Glottolog currently lists more than 5,000 unique languages, such an undertaking is beyond the scope of this manuscript. Achieving this goal would require substantial community collaboration, together with sustained feedback and guidance from domain experts specialising in particular regions and language families. As a first step in this direction, we encourage the research community to provide feedback via GitHub’s Issues mechanism, in the form of objections or comments on specific speaker areas, their geometry and metadata.

Language areas are constructed spaces that indicate the presence of a language in a specific geographic region. They are derived from the presence of individuals speaking, signing, or writing a particular language, making them a useful simplification of linguistic reality. One could also imagine that individual utterances tagged with a language and a location, or the prolonged presence of a speaker producing them, could directly signal the presence of a language in a region. While this approach stays closer to discretely measurable observations, it comes at the expense of feasibility, simplicity, and interpretability, and will likely remain an aspirational goal.

5.2 Conclusion

Glottography is a free and open repository providing digital polygons for the world’s languages. Currently, it contains over 13,000 language polygons representing more than 5,300 unique languages. The coverage of Glottography largely matches that of the only comparable service, the subscription-based World Mapping System offered by Ethnologue. All polygons are available in the Cross-Linguistic Data Format on Zenodo, ready for use in Geographic Information Systems or any spatial programming environment, e.g. through the Rglottography package in R. Detailed tutorials encourage community members to contribute their own language polygons to the platform, expanding coverage to previously uncharted regions, languages, and historical epochs.