CultureCartography: Mapping Cultural Structures
- CultureCartography is the computational mapping of cultural phenomena, using geospatial, network, and semantic methods to reveal cultural structures.
- It integrates cartographic design, audience-centric web mapping, and semantic analysis to quantify commemorative practices, textual shifts, and knowledge spaces.
- The approach operationalizes cultural data with interoperable models and mixed-initiative systems, enabling reproducible research across urban, digital, and historical domains.
CultureCartography is the cartographic and computational treatment of culture as a mappable domain. In its narrowest current usage, “Culture Cartography” is a mixed-initiative methodology for eliciting culture-specific knowledge that is salient to in-group users and missing from LLMs (Ziems et al., 31 Oct 2025). A broader reading of recent research suggests a wider program: commemorative street names can be mapped as public acts of remembrance, web use can be mapped as geo-linguistic audience regions, texts can be mapped as semantic spaces, collections can be mapped as navigable knowledge terrains, historical cities can be reconstructed as spatiotemporal references, and maps themselves can be studied as cultural artifacts shaped by institutions, semiotic conventions, and political histories (Bogucka et al., 2021, Taneja, 2016, Stoltz et al., 2020, Scharnhorst, 2015, Tavakkol et al., 2020, Petitpierre, 24 Nov 2025). Across these variants, maps are treated not merely as navigation devices but as selective, rhetorical, and machine-processable representations of cultural structure.
1. Conceptual foundations
CultureCartography rests on the proposition that maps are culturally meaningful objects and that cultural phenomena can be rendered through spatial, graph, and semantic form. In urban cartography, street names are treated as commemorative spaces rather than arbitrary labels, so a city’s naming system can reveal who is celebrated, who is ignored, which professions are honored, how gender bias manifests, and whether the city symbolically opens itself to foreign cultures (Bogucka et al., 2021). In web studies, infrastructure-centered maps based on hyperlinks are contrasted with audience-centric maps based on shared traffic, with the latter argued to reflect cultural proximity, especially language and geography, more directly (Taneja, 2016, Wu et al., 2015). In computational text analysis, word embeddings are proposed as a form of cultural cartography because they preserve graded and relational meaning better than keyword counts (Stoltz et al., 2020).
The same conceptual shift appears in work on libraries, archives, and scholarly corpora. Knowledge maps are framed as “macroscopes” that render the “structure and evolution of data, information and knowledge” and support navigation “across the lands and oceans of knowledge” (Scharnhorst, 2015). Science mapping extends this logic by treating circulation as something that can be visualized through co-authorship, bibliographic coupling, and co-word networks, thereby exposing centers, peripheries, bridges, and research fronts (Cortes et al., 2021). Historical cartography at scale pushes further by arguing that maps are semantic-symbolic systems and cultural objects reflecting political and epistemic expectations, not transparent recordings of territory (Petitpierre, 24 Nov 2025).
This literature also rejects the idea that cartographic output is a neutral consequence of data. The multirepresentation study of world population states that “Toute réalisation cartographique est affaire de choix” and demonstrates, with thirteen different maps from one dataset, that cartography is governed by “l’intention cartographique” rather than a mechanical relation between data and graphic form (Lambert et al., 2023). Work on collaborative map vandalism strengthens the same point from another angle: collaborative maps are “loaded with cultural, aesthetic, and practical meanings,” and their defacement reveals that cartographic artifacts are contested surfaces within the digital commons (Ballatore, 2014).
2. Principal objects and data models
The objects of CultureCartography vary by domain, but they are consistently normalized into relational or geospatial schemas that support comparison, filtering, and composition. In commemorative street-name mapping, the core record includes street name, district, denomination date, honoree, gender, occupation, country of origin, date of birth, and date of death; the associated cross-city corpus contains 4,932 curated honorific streets across Paris, Vienna, London, and New York (Bogucka et al., 2021). In audience-centric web mapping, the basic entities are websites or subdomains, and ties are built from shared audience traffic among the top 1000 globally popular domains, operationalized on 973 websites in one study and around 1018, 1022, and 1031 websites across 2009, 2011, and 2013 in another (Taneja, 2016, Wu et al., 2015). In semantic cartography, the raw structure is the term-context matrix or document-term matrix, reduced into embedding spaces, semantic directions, or transport distances such as Word Mover’s Distance and Concept Mover’s Distance (Stoltz et al., 2020).
In collection and scholarship mapping, the objects are books, articles, authors, institutions, categories, classifications, and key-terms. KnoweScape explicitly lists library, archival, and museum collections; national bibliographies; Wikipedia; social networks; collaborative platforms such as Mendeley and CiteULike; scholarly communication databases; research information systems; and funding databases as candidate knowledge spaces (Scharnhorst, 2015). Science mapping treats institutions, documents, and keywords as nodes in co-authorship, bibliographic coupling, and co-word networks (Cortes et al., 2021). Historical map scholarship scales this further: ADHOC Records aggregates 771,561 cleaned map records from 38 digital catalogs, while ADHOC Images contains 99,715 digitized map images after normalization and filtering (Petitpierre, 24 Nov 2025).
The same plurality appears in digital systems and benchmarks. Kartta Labs uses scanned maps, building footprints, roads, addresses, names, dates, start dates, end dates, historical photographs, annotated facades, and 3D models as linked spatiotemporal resources (Tavakkol et al., 2020). CartoMark assembles around 20,000 original maps harvested from search engines, repositories, and some social media, then labels them for map text annotation recognition, map scene classification, super-resolution reconstruction, and style transferring; its scene schema spans map dimension, map theme, map style, map view, and map hue (Zhou et al., 2023). The LLM-oriented version of Culture Cartography collects editable question-answer trees about cultural practices, with data gathered in Nigeria and Indonesia from annotators spanning 7 and 13 ethnolinguistic groups, respectively (Ziems et al., 31 Oct 2025).
| Modality | Primary objects | Representative work |
|---|---|---|
| Commemorative urban mapping | Streets, honorees, districts, dates | (Bogucka et al., 2021) |
| Audience-centric web mapping | Domains, shared audiences, clusters | (Taneja, 2016, Wu et al., 2015) |
| Semantic text mapping | Words, documents, embeddings, semantic directions | (Stoltz et al., 2020) |
| Knowledge-space mapping | Collections, classifications, keywords, institutions | (Scharnhorst, 2015, Cortes et al., 2021) |
| Historical urban reconstruction | Maps, footprints, facades, temporal intervals, 3D meshes | (Tavakkol et al., 2020) |
| Map-at-scale analysis | Records, images, signs, fragments, contributors | (Petitpierre, 24 Nov 2025, Zhou et al., 2023) |
| Mixed-initiative cultural elicitation | Seed topics, editable questions, answers, preference pairs | (Ziems et al., 31 Oct 2025) |
A plausible implication is that CultureCartography is less a single data standard than a family of interoperable abstractions: geospatial features, networks, semantic vectors, and annotated visual corpora are all used as carriers of cultural structure.
3. Methods of representation and analysis
One major branch of CultureCartography uses explicitly cartographic techniques. “Cartographic Design of Cultural Maps” organizes cultural street-name exploration through a three-arc narrative sequence—setup, conflict, resolution—and five interface sections: a pre-loader, a hook, a conflict stage, the interactive map, and a denouement (Bogucka et al., 2021). It deploys pointillism, scale-dependent “zoomy-telling,” strong figure-ground contrast, person-centered popups, temporal filtering, category filtering, and a “Random street” interaction to support serendipitous discovery. The multirepresentation chapter demonstrates a complementary principle: one dataset can be mapped through choropleth, proportional symbols, cartograms, dot maps, smoothed potential surfaces, network thresholds, 3D extrusions, and thresholded void maps, each foregrounding a different spatial reading (Lambert et al., 2023).
A second branch is network-analytic. Audience-centric web maps are built by defining a tie only when observed audience duplication exceeds expected duplication under independence. One formulation is
with valued ties representing greater-than-expected duplication (Taneja, 2016). The longitudinal ethnological mapping of the web uses above-random duplication together with force-directed layouts, hierarchical clustering, distance, and the E-I index,
to characterize the size, distance, and “thickness” of online regional cultures (Wu et al., 2015). Science mapping uses density, average path length, modularity, and betweenness centrality across co-authorship, bibliographic coupling, and co-word networks to identify institutional brokerage, research fronts, and knowledge pillars (Cortes et al., 2021). The “Web Maps and Their Algebra” paper abstracts this still further by defining maps of graphs as connectivity summaries over distinguished nodes and equipping them with operations analogous to join and meet, making map composition formally machine-processable (Fionda et al., 2013).
A third branch is semantic. Cultural cartography with word embeddings begins from the term-context matrix and moves into lower-dimensional vector spaces where similarity is computed geometrically. The paper relies primarily on cosine similarity,
semantic directions derived from averaged vector differences, and two complementary navigation strategies: variable embedding space, which holds terms constant and studies semantic change over time or across fields, and fixed embedding space, which holds the semantic field constant and studies the movement of documents or organizations relative to it (Stoltz et al., 2020). This makes symbolic boundaries, social marking, echo chambers, and cultural diffusion tractable as distances, projections, and transport costs rather than as isolated word counts.
The LLM-oriented methodology named “Culture Cartography” adds a mixed-initiative elicitation procedure. An LLM proposes up to 5 questions and up to 5 answers around an editable seed topic; humans can edit, delete, regenerate, or author nodes directly; and low-confidence answers are highlighted by prompting the same model with “Does this answer the question correctly?”, constraining logits to True/False, and marking answers uncertain when the probability of True is (Ziems et al., 31 Oct 2025). Here the “map” is a branching knowledge tree rather than a geographic surface, but the underlying logic is consistent with other forms of CultureCartography: culture is explored through structure-preserving reduction, selective emphasis, and navigable relations.
4. Infrastructures, interfaces, and operational systems
CultureCartography is also a systems problem. The cultural street-name platform stores geodata in PostgreSQL, matches each street to its OpenStreetMap shapefile, and renders the interface with Mapbox GL JS (Bogucka et al., 2021). Kartta Labs is explicitly modular and open-source, with Maps and 3D models as major modules and sub-modules such as Warper, Editor, Server, Kartta, Noter, Parser, Reconstructor, Reservoir, and Renderer (Tavakkol et al., 2020). Its temporal model is interval-based through start date and end date attributes, enabling client-side filtering of all features in a tile according to a selected date. The platform couples crowdsourcing and AI: users georeference historical maps, trace features, annotate facades, and link photos to footprints, while AI supports initial geolocation guesses, facade parsing with binary Faster R-CNN, and related recognition tasks.
Knowledge-space projects emphasize interoperability rather than a single interface stack. KnoweScape foregrounds crosswalks among classifications, taxonomies, thesauri, topic maps, ontologies, linked data representations, and SKOS-like infrastructures, because machine-harvestable interoperability is itself a prerequisite for mapping collections (Scharnhorst, 2015). The graph-cartographic formalism of “Web Maps and Their Algebra” explicitly envisions maps as machine-readable objects, for example in RDF, that can be shared, exchanged, reused, and composed (Fionda et al., 2013). CartoMark contributes benchmark infrastructure for map-specific machine learning by packaging text annotations, scene classes, super-resolution pairs, and style corpora in standardized file formats, with duplicate filtering based on SSIM and NMI and manual consensus procedures for scene labels (Zhou et al., 2023).
Reproducibility is a parallel infrastructural theme. The multirepresentation chapter executes acquisition, treatment, and cartographic rendering mainly in R with packages including raster, sf, cartography, SpatialPosition, tanaka, and linemap, while acknowledging that final page composition in Adobe Illustrator breaks strict reproducibility (Lambert et al., 2023). The LLM-focused CultureExplorer tool is implemented on the Farsight codebase and turns mixed-initiative elicitation into a web interface centered on a tree of questions and answers rather than a linear chat (Ziems et al., 31 Oct 2025). This suggests that CultureCartography increasingly depends on dual infrastructures: analytic back ends for computation and data normalization, and exploratory interfaces that preserve traceability while supporting public or expert interpretation.
5. Empirical domains and findings
The empirical reach of CultureCartography is unusually broad. In urban commemoration, the cultural street-name maps show spatially localized gender bias and policy response: Vienna exhibits clusters of streets named after female figures in the last decade, particularly near Seestadt Aspern; similar newly formed female clusters are reported in south Paris (13th arrondissement) and Brooklyn (Bogucka et al., 2021). Occupational commemoration differs sharply across cities: London heavily commemorates royals and politicians, Vienna emphasizes artists and writers, New York celebrates social activists and lacks military professionals in the dataset, and Paris commemorates generals and soldiers associated with the French Revolution, Napoleonic wars, and the World Wars. For foreign cultures, Vienna shows a broader range of national origins than Paris, London, or New York.
In web cartography, the audience-centric graph and the hyperlink graph produce markedly different structures. On a comparable 2012 sample, the hyperlink network has density = 0.048, clustering coefficient = 0.524, and network centralization = 68%, whereas the audience network has density = 0.395, clustering coefficient = 0.846, and network centralization = 52% (Taneja, 2016). The audience graph reveals 9 communities aligned largely with geo-linguistic similarity, and the QAP correlation between valued hyperlink ties and valued audience-overlap ties is only , indicating that technical linking and shared use are largely different structures. The longitudinal ethnological mapping extends this argument: network density declines from 0.43 in 2009 to 0.35 in 2013, while clustering coefficient stays around 0.85–0.86, and clusters associated with China, India, Brazil, and the former USSR/Russia grow and thicken, interpreted as the emergence of more regionally grounded online cultures (Wu et al., 2015).
In semantic discourse analysis, the immigration case study shows long-run changes in meaning space rather than only lexical frequency. The cosine similarity between “immigration” and “crime” is lowest in the 1890s at 0.076 and peaks in the 1990s at 0.344, while the next closest term in that decade is “school” at 0.087 (Stoltz et al., 2020). Fixed-space document analysis further finds that right-leaning outlets become more semantically similar to right-wing immigration advocacy organizations than to left-wing ones when writing about immigration, whereas left-leaning outlets are roughly equally similar to both sides. The same paper shows co-movement between immigrant/citizen and black/white semantic dimensions in U.S. news, interpreted as evidence of persistent racialized marking.
Studies of map corpora reveal that the objects of CultureCartography are themselves historically patterned. The multilingual analysis of 23,928 academic thematic maps finds that Chinese- and English-language journals share highly similar conventions—neutral dominant hues, low saturation, high brightness, limited hue diversity, centered layouts, and high main-map occupation ratios—while differences such as slightly greater hue richness and compactness in English-language maps remain small in magnitude (Wei et al., 24 Apr 2026). The large-scale history of cartographic heritage finds a domestic-focus trend that rises from roughly 40% in the late eighteenth century to nearly 70% in the late nineteenth century and reports a Pearson correlation above 0.91, peaking at with an eight-year lag, between Atlantic maritime charting and enslaved-captive embarkation counts (Petitpierre, 24 Nov 2025). Such results move CultureCartography beyond localized case studies toward quantified historical patterning.
The LLM-oriented methodology adds an evaluative empirical result: CultureExplorer produces cultural data that strong models still miss. On Culture Cartography data, DeepSeek R1 reaches recall 0.85 for Indonesia and 0.82 for Nigeria, lower than on Traditional Annotation data; fine-tuning Llama-3.1-8B on this data raises performance by up to 19.2% on related culture benchmarks (Ziems et al., 31 Oct 2025). This suggests that CultureCartography can function not only as analysis and visualization, but also as a method for discovering long-tail cultural knowledge not easily captured by pretraining or web search.
6. Institutions, contestation, and limits
CultureCartography is shaped by institutional conditions as much as by methods. The interview study of 16 cartographers and GIS experts from 13 organizations shows that choropleth production follows a five-stage workflow—data preparation, data analysis, data binning, map styling, and post-processing—but that actual decisions are negotiated under brand rules, review chains, legal constraints, accessibility requirements, recurrent reporting conventions, and client pressure (Narechania et al., 13 Apr 2025). The most common classification methods are pretty breaks and manual interval, not because they are universally optimal, but because they are readable, explainable, and organizationally sustainable. This reframes cartography as situated work rather than pure visual optimization.
Collaborative mapping introduces another institutional layer: governance of the digital commons. Carto-vandalism is defined as intentional defacement that reduces the utility of a geospatial artifact for the majority of users and is classified into play, ideological, fantasy, artistic, industrial carto-vandalism, and carto-spam (Ballatore, 2014). The significance for CultureCartography is twofold. First, maps are culturally charged surfaces that attract symbolic struggle, parody, commercial capture, and aesthetic intervention. Second, open cultural maps require policing, moderation, and automated detection, so their authority depends on ongoing community labor and norm enforcement.
Several methodological limits recur across the literature. The cultural street-name platform offers design rationale and scenario-based demonstration but no formal user study, controlled experiment, or quantitative usability assessment (Bogucka et al., 2021). Audience-centric web mapping depends on domain-level panel data and incomplete crawling, making it difficult to disaggregate global platforms or attribute co-use purely to culture rather than search, recommendation, or third-party pathways (Taneja, 2016, Wu et al., 2015). Embedding-based cultural cartography requires alignment across spaces, careful corpus selection, and theoretical interpretation because embeddings do not “understand” meaning (Stoltz et al., 2020). Historical reconstruction platforms such as Kartta Labs accept incompleteness and lack formal uncertainty models, confidence fields, or alternative reconstructions when archival evidence is ambiguous (Tavakkol et al., 2020). Large benchmark datasets such as CartoMark are strong on visual heterogeneity but weak on temporal labels, cultural provenance, and explicit retrieval protocols (Zhou et al., 2023).
A broader interpretive limit is that “culture” itself is not a single scale or ontology. Some work maps commemorative practices, some maps co-attention, some maps semantic fields, some maps collections, and some maps the internal culture of cartographic production. This suggests that CultureCartography is best understood as a plural research formation rather than a closed theory. Its unifying principle is not one method or one object, but the systematic transformation of cultural relations into explicit structures—geometric, topological, semantic, or interactive—that can be explored, compared, and contested.