PaperlayerView exact revision on arXiv

Source-linked AI summary

Terrain signatures in Welsh settlement names

Oktay Karakuş, Can Eyupoglu

arXiv:2608.26978v1cs.LGcs.CL

TL;DR

The paper examines whether Welsh settlement-name categories retain measurable environmental information beyond broad geographic structure. It applies a frozen lexical framework, preregistered outcome-specific models, and geographically structured validation, finding terrain-name information in present-day topographic position and held-out prediction. The conclusion is bounded to selected categories within Wales and does not establish individual etymology, causation, historical environmental memory, or transferability.

  • Problem

    The study addresses whether place names retain measurable environmental information beyond broad geographic structure, because their relationship to contemporary conditions must be tested rather than assumed.

  • Method

    The analysis uses a frozen, source-audited lexical framework with preregistered outcome-specific models and geographically structured validation of Welsh settlement-name categories.

  • Results

    24.4 m higher relative topographic position was observed for high-terrain names versus low-terrain names, while terrain-name polarity reduced geographically held-out mean squared error by 4.63%, 6.22%, and 7.30% under 10-, 25-, and 50-km blocking.

  • Takeaways & Limitations

    Selected Welsh settlement-name categories retain measurable information about present-day terrain within Wales beyond the stated geographic and settlement baseline.

  • Takeaways & Limitations

    Predictive improvement varied among held-out regions, so terrain-name information should be interpreted as an aggregate out-of-area signal rather than a geographically uniform one.

Abstract

from arXiv · show

Landscapes are named, but whether names retain measurable environmental information beyond broad geographic structure is rarely tested. We analysed 3,757 Welsh settlements using a frozen, source-audited 24-element lexical framework, preregistered outcome-specific models and geographically structured validation. The central comparison contrasted 101 settlements carrying high-terrain elements (\textit{bryn} or \textit{mynydd}) with 139 carrying low-terrain elements (\textit{cwm} or \textit{pant}). High-terrain names occupied locations 24.4 m higher relative to their 2-km surroundings (95\% CI, 10.8--38.1 m; Holm-adjusted $p$ = 0.00137). The association remained positive across prespecified 1-, 2- and 5-km neighbourhood definitions and was reproduced using an independently produced elevation source (24.1 m; 95\% CI, 10.6--37.6 m). Adding terrain-name polarity to a non-lexical spatial and settlement baseline reduced geographically held-out mean squared error by 4.63\%, 6.22\% and 7.30\% under 10-, 25- and 50-km spatial blocking, respectively, although improvement varied among held-out regions. River-related names provided weaker, directionally consistent evidence, while the preregistered woodland model was non-estimable. Residual spatial structure, unresolved name language and the absence of independent external replication limit interpretation. Selected Welsh settlement-name categories therefore retain measurable information about present-day terrain within Wales, without establishing individual etymology, causal naming, historical environmental memory or transferability to other naming systems.

Introduction

The paper tests whether reproducibly classified Welsh place-name categories retain measurable information about present-day environmental conditions beyond geographic and settlement baselines. It combines preregistered association analyses with spatially structured validation while treating lexical categories as exposures rather than individual etymological interpretations.

  • Place names can reflect terrain, water, vegetation, land use, ownership, memory, language, and administrative histories, so environmental correspondence must be tested rather than assumed.
  • The study asks whether a predeclared, source-attributed Welsh lexical classification contains measurable information about specified present-day environmental variables at settlement locations.
  • The lexical screen is a reproducible exposure definition, not a validated resolution of every name’s morphology, language status, historical sense, or environmental reference.
  • Spatially structured validation is required because nearby settlements may share geographic context, while residual spatial dependence and observational associations remain interpretive limitations.
  • The analysis uses a Wales-wide settlement frame, a frozen rule-based screen, and a source-audited registry of 24 Welsh place-name elements.
  • Three outcome-specific hypotheses concerned river proximity, terrain polarity, and woody cover, with preregistered models, sensitivity analyses, spatially restricted randomisation, residual diagnostics, and geographically held-out prediction.

Results

Welsh settlements with high-terrain name elements occupied higher local topographic positions than settlements with low-terrain elements, and this association was robust across analytical choices. Terrain-name polarity also improved geographically held-out prediction beyond a non-lexical spatial and settlement baseline, although regional improvements were uneven.

  • Terrain-name contrast: Local topographic position measured settlement elevation relative to surrounding terrain within a 2-km neighbourhood, rather than absolute elevation.Positive and negative values distinguish locally elevated and depressed positions.
  • Terrain-name contrast: 24.4 m higher local topographic position was observed for high-terrain than low-terrain settlements after adjustment (95% CI: 10.8–38.1 m; Holm-adjusted p = 0.00137).The comparison included 101 high-terrain settlements containing bryn or mynydd and 139 low-terrain settlements containing cwm or pant.
  • Robustness: 19.2 m at 1 km and 25.9 m at 5 km preserved the positive high-minus-low contrast, indicating robustness to neighbourhood scale.The 2-km estimate was the preregistered primary analysis.
  • Robustness: 24.1 m using an independent Copernicus elevation product closely matched the 24.4-m primary estimate.Additional checks using one record per settlement name and inverse name-group-size weighting also produced stable estimates, including 21.3 m and 23.6 m.
  • A Geographically held-out prediction: 4.63%, 6.22% and 7.30% reductions in pooled held-out MSE occurred when terrain-name polarity was added under 10-, 25- and 50-km spatial blocking.The improvement occurred across all three blocking scales but was not uniform among held-out regions.
  • Other lexical hypotheses: River-related names showed a weaker directional association with proximity to mapped river networks, while the preregistered woodland model was non-estimable.The river analysis estimated a multiplicative ratio of 0.649 for 1 + river distance; the woodland model did not converge and yielded no interpretable estimate.

Discussion

The preregistered terrain-name contrast corresponds to measurable differences in local topographic position, with evidence remaining positive across sensitivity analyses and geographically held-out prediction. The findings support a bounded Welsh result rather than claims about individual etymology, causation, historical environmental memory, or broader transferability.

  • Terrain association: 24.4 m higher relative to 2-km surrounding terrain was the estimated difference for settlements containing bryn or mynydd versus cwm or pant.The direction persisted across 1-, 2- and 5-km neighbourhoods and an independently produced elevation source.
  • Predictive validation: 4.63%, 6.22%, and 7.30% reductions in held-out mean squared error occurred under 10-, 25-, and 50-km spatial blocking after adding terrain-name polarity to the baseline.The contribution was driven primarily by the prespecified terrain-polarity distinction, not lexical information in general.
  • Predictive validation: Regional improvement was heterogeneous, occurring in three of five folds at 10 km, three of five at 25 km, and four of five at 50 km.The increasing pooled percentages do not establish a scale-response relationship or increasing informativeness with geographic distance.
  • Interpretive boundary: The contribution is a reproducible framework testing frozen lexical exposures against contemporary environmental conditions beyond a stated geographic and settlement baseline.It does not reconstruct individual etymologies or establish when names arose, what motivated them, or historical environmental memory.
  • Scope across lexical domains: River-related names supplied weaker directional evidence, whereas the preregistered woodland model did not converge and yielded no interpretable confirmatory estimate.These secondary results do not support equal recoverability of environmental information across lexical domains.
  • Limitations: Residual spatial autocorrelation, unresolved Welsh-language labels, and a non-externally validated lexical screen constrain interpretation of the associations and predictions.The terrain-source comparison also does not establish interchangeability, and external replication beyond Wales remains necessary for broader generality.

Conclusions

Within Wales, selected terrain-name categories retain measurable information about present-day local topography beyond fitted spatial and settlement baselines. The evidence is strongest for terrain polarity, while several limitations bound interpretation.

  • Selected Welsh terrain-name categories retain measurable information about present-day local topography beyond the fitted spatial and settlement baseline.
  • Settlements containing bryn or mynydd occupied systematically higher local positions than settlements containing cwm or pant.
  • The terrain contrast persisted across neighbourhood definitions, an independent elevation source, and analyses reducing repeated-name dependence.
  • Terrain-name polarity improved prediction for geographically held-out settlements under 10-, 25- and 50-km spatial-blocking designs.
  • River-related names showed weaker evidence, whereas the preregistered woodland model was non-estimable.
  • Interpretation is limited by regional variation in predictive gains, residual spatial structure, unresolved name language, and the lack of independent external replication.

Methods

The study used a frozen, source-audited lexical screen to define Welsh settlement-name exposures before environmental analyses. Deterministic token matching supported preregistered terrain, water, and woodland contrasts without resolving individual etymology.

  • The Wales-wide frame retained 3,757 settlement records represented as point locations, with repeated names kept as separate locations.
  • A source-audited registry of 24 Welsh place-name elements defined the lexical exposures before outcome analysis.
  • Names were lowercased, normalised, tokenised, and assigned an element when a token began with that element.
  • The deterministic screen retained exact-token and broader-prefix matches separately for prespecified sensitivity analyses.
  • The screening rule was not a morphological parser, language classifier, or semantic disambiguation model and did not validate individual-name etymology.
  • H1 contrasted afon or nant matches with other eligible settlements, H2 contrasted bryn or mynydd with cwm or pant, and H3 contrasted coed or llwyn with others.

Environmental outcomes

Environmental outcomes operationalised river proximity, local terrain position, and woodland cover using independently defined spatial data and outcome-specific models. Confirmatory inference used covariate adjustment, clustered uncertainty, and Holm correction, but woodland was non-estimable.

  • H1 measured log(1 + river distance), H2 measured local terrain position, and H3 measured 1-km woodland cover.
  • H2 terrain position was point elevation minus mean elevation within a circular 2-km buffer, with 1-km and 5-km sensitivity values also calculated.
  • Buffered environmental values were withheld below 95% valid raster-cell or land-cover coverage, and no numeric covariates were imputed.
  • Models adjusted for settlement characteristics, name attributes, non-target lexical detection, and a spatial spline of British National Grid coordinates.
  • The preregistered H3 fractional-logit model failed to converge, so no alternative feasibility model was substituted.
  • H2 coefficients were reported as high-terrain minus low-terrain differences in metres, with two-sided 95% confidence intervals and Holm-adjusted p-values.

Spatially restricted randomisation

The study tested terrain-name robustness through geographically restricted reassignment and spatially blocked prediction. These analyses preserved local structure while evaluating whether terrain polarity added information beyond a non-lexical baseline.

  • Restricted null analysis generated 1,999 two-sided permutations by shuffling H2 labels within geographic, settlement, language-status, and name-length strata.
  • The restricted randomisation preserved environmental outcomes, covariates, model specification, and stratum membership.
  • The reassignment test preserved specified local spatial and settlement structure but did not imply complete spatial exchangeability or remove residual dependence.
  • Separate five-fold spatial partitions evaluated prediction under 10-, 25-, and 50-km blocking designs.
  • The non-lexical baseline M0 used settlement-structure covariates and a spatial spline, while M2 added the high-versus-low terrain-polarity indicator.
  • Held-out performance used pooled mean squared error, with penalty selection confined to outer-training data.
  • The blocked analyses quantified geographically held-out prediction within Wales rather than external validation or a continuous distance-response relationship.

Terrain-source and dependence robustness

Robustness analyses preserved the primary terrain-name contrast across neighbourhood definitions, alternative elevation data, and dependence checks. The independent elevation source closely reproduced the primary estimate, while coordinate-cluster deduplication left the result unchanged.

  • Neighbourhood robustness: The local topographic-position outcome was recalculated using 1- and 5-km neighbourhoods alongside the preregistered 2-km definition.These changes preserved the primary terrain-name contrast.
  • Interpretive boundary: The elevation-source comparison was conducted post-registration and assessed robustness to an alternative product rather than source interchangeability.The primary OS Terrain 50 product and Copernicus DEM GLO-90 differ in data-product type.
  • Dependence robustness: Coordinate-cluster deduplication removed no observations because all 240 H2 observations occupied distinct coordinate clusters.The procedure reproduced the primary result exactly.

Data availability

Publication-safe numerical source data and related documentation are available through an archived reproducibility release, while raw third-party spatial data are not redistributed.

  • Source data: Publication-safe source data, including numerical data underlying the main figures and tables, are available in the archived reproducibility release.The release is identified by DOI 10.5281/zenodo.22079109.
  • Third-party data: Raw third-party spatial data are not redistributed; source datasets, access routes, and licensing conditions are documented in the accompanying repository.

Code availability

Custom analysis code and reproducibility materials are publicly available in a frozen, permanently archived scientific release with claim-traceability and integrity records.

  • Code: Custom analysis code and reproducibility materials supporting the reported results are publicly available.The materials are hosted in the stated GitHub repository.
  • Archived release: The frozen scientific release is version v1.0.0 and is permanently archived at DOI 10.5281/zenodo.22079109.
  • Release contents: The archived release includes claim-bearing analysis code, frozen aggregate outputs, source-data materials, a release manifest, claim-traceability records, and SHA-256 checksums.

Supplementary information

The Supplementary Information provides detailed methodological, diagnostic, validation, and reproducibility material, while claim-essential estimates remain in the main text.

  • Supplementary scope: The Supplementary Information covers study-frame and lexical provenance, environmental outcome construction, confirmatory and spatial-analysis methods, robustness analyses, diagnostics, and held-out validation.
  • Supplementary tables: Supplementary Tables S1–S7 provide the lexical inventory, analytical populations, sensitivity analyses, diagnostics, spatial-CV decomposition, and complete fold-level results.
  • Role of the Supplement: Claim-essential estimates remain in the main text, while the Supplement provides audit and reproducibility detail rather than a second narrative.
Loading 2608.26978v1…