PaperlayerView exact revision on arXiv

Source-linked AI summary

Where meanings converge and diverge: Revisiting linguistic relativity across 81 languages

Weiran Wang, Wenjuan Qin, Haitao Liu

arXiv:2610.30778v1cs.CL

TL;DR

The paper asks whether speakers of different languages organize basic concepts in similar ways, despite shared meanings not guaranteeing shared relations. It compares language-derived relational structures across 196 concepts and 81 languages, finding greater correspondence for sampled Socio-Cultural than Nature & Motion meanings, alongside domain-specific geographic and broader genealogical associations.

Abstract

from arXiv · show

Do people who speak different languages organize concepts in similar ways? This question lies at the heart of the Sapir--Whorf hypothesis and debates about linguistic relativity. Shared basic meanings, however, need not imply shared relations among them. We compared relations among 196 basic concepts across 81 languages, using word embeddings from two multilingual models to quantify conceptual structure expressed in language. Cross-language correspondence was greater for sociocultural meanings than for meanings concerning nature and motion, a contrast recurring across models and classifications. Within the sociocultural domain, relational differences increased with geographic separation; less closely related languages differed more across all domains. These findings suggest that universality and relativity are interwoven across conceptual domains. They inform an account in which shared cognitive resources interact with linguistic and local experience, and predict greater cross-language correspondence in human semantic judgments for sociocultural than for nature and motion meanings.

Introduction

The introduction frames linguistic relativity as a question about how shared cognitive resources and linguistic experience shape conceptual organization. It motivates comparing relations among meanings across domains and outlines a computational comparison of 196 concepts across 81 languages.

Introduction

  • Translation equivalents can refer to corresponding things yet occupy different semantic neighborhoods, making cross-language relations important to conceptual coherence, categorization, and generalization.Source: Introduction · p. 1

Languages, Concepts, and Semantic Domains

  • The study compares translation-equivalent concepts across languages using a coverage-based sample, retaining common support of 81 languages and 196 concepts in both encoders.The starting inventory covered 82 languages and 207 concepts; the retained support followed exclusions for unusable representations.Source: Languages, Concepts, and Semantic… · p. 3
  • Four author-defined domains group the concepts, but they are operational groupings rather than independently established conceptual modules.A reference-style grouping reassigned 22 input concepts; the two groupings agreed on 185 input concepts, producing a 175-concept consensus subset.Source: Languages, Concepts, and Semantic… · p. 3, Table 1 · p. 3

Lexical Representations and Relational Dissimilarity

  • Two multilingual encoders provide lexical representations, and representational similarity analysis compares the ordering of semantic relations within each language.The encoders use the same forms, pooling rules, and support while retaining their own tokenizers and training histories.Source: Lexical Representations and Relational… · p. 3

Geographic and Linguistic Distance Measures

  • Geographic separation and estimated genealogical distance are entered jointly, so each distance slope describes its association conditional on the other.The genealogical proxy can reflect both inheritance and borrowing; geographic separation does not directly measure contact.Source: Geographic and Linguistic Distance… · p. 4

Statistical Analysis and Robustness Checks

  • Domain-specific Bayesian regressions estimate adjusted dissimilarities and distance slopes, with shared language intercepts accounting for repeated participation in language pairs.Robustness analyses regrouped concepts, used a consensus subset, permuted concept labels, and widened prior scales; secondary frequency and archived-vector checks were not matched replications of the primary RSA analysis.Source: Statistical Analysis and Robustness… · p. 4

Results

Across the sampled languages, Socio-Cultural relations corresponded more closely than Nature & Motion relations, while geographic and genealogical distance showed different associations with relational dissimilarity.

Semantic Similarity Across Languages Varies by Domain

  • The Socio-Cultural–Nature & Motion contrast survived regrouping and consensus-concept analyses, though Grammar & Function became lowest-dissimilarity in the mBERT consensus subset.All 16 encoder-by-grouping-by-domain outcomes fell below all 2,000 label permutations (Monte Carlo lower-tail p< .001).Source: Semantic Similarity Across Languages… · p. 5
  • Socio-Cultural dissimilarity was lower than Nature & Motion in both mBERT (.874 vs .924) and XLM-R-large (.922 vs .977).The domain contrast persisted across encoders, although the complete profiles differed.Source: Figure 1 · p. 5, Semantic Similarity Across Languages… · p. 5, Figure 2 · p. 6

Semantic Differences Are Associated With Geographic Distance and Language Relatedness

Illustrative Comparisons Among Individual Languages

  • Selected triplets illustrated distinct distance configurations: Dutch–Afrikaans often exceeded Dutch–Basque, while Scots exceeded Marathi on selected Socio-Cultural items.These selected item-level examples do not independently test the domain-level RSA results or identify contact and inheritance as causes.Source: Illustrative Comparisons Among… · p. 7, Figure 4 · p. 8

Discussion

The discussion frames cross-language conceptual organization as shared at some levels and variable at others. It interprets the domain and distance patterns through shared cognitive resources, experience, and social use, while identifying behavioral validation as necessary.

Shared Meanings, Different Relations

  • Relations among sampled Socio-Cultural meanings corresponded more closely across languages than Nature & Motion relations, a contrast recurring across encoders and classification changes.The profile refines the linguistic-relativity question by locating variation within relations among basic meanings.Source: Shared Meanings, Different Relations · p. 9

Shared Constraints and Local Experience

  • Shared relational constraints and recurrent communicative demands may favor similar patterns, while locally varying experience may favor different distinctions.The discussion connects the contrast to internal structure, communicative efficiency, and ecological experience.Source: Shared Constraints and Local Experience · p. 9
  • Socio-Cultural relations were comparatively similar overall yet more dissimilar with geographic separation, while greater genealogical distance tracked dissimilarity across domains.The discussion presents average correspondence and geographic gradients as distinct properties that can coexist.Source: Shared Constraints and Local Experience · pp. 9–10

Universality and Relativity in an Adaptive System

  • The authors frame universality and relativity as interwoven properties of adaptive meaning systems and predict greater cross-language correspondence in human judgments for sampled Socio-Cultural than Nature & Motion relations.They propose comparisons with nonverbal materials and manipulations of language use as further tests.Source: Universality and Relativity in an… · p. 10
  • The evidence is model-derived lexical relations, so correspondence with speakers’ judgments remains to be established; training biases and sampling also constrain generalization.Communicative, ecological, and historical mechanisms require independent measurements.Source: Universality and Relativity in an… · p. 10

S1 Sample construction and quality control

The study assembled a multilingual inventory of lexical forms and concept vectors, then applied quality checks to define the retained language and concept support.

S1 Sample construction and quality control

  • The starting inventory comprised 82 languages and 207 concepts, drawn from multilingual BERT coverage and Wiktionary’s language-specific Swadesh lists.The 82-language, 207-concept inventory was inherited from data assembly rather than introduced by the quality-control procedure.Source: S1 Sample construction and quality… · p. 14
  • Quality checks identified 12 all-zero observations, and Javanese was excluded because its maximum duplicate share was .607, below the archived file’s .995.Javanese remained degenerate after extraction and was excluded under the stated duplicate-share rule.Source: S1 Sample construction and quality… · p. 14
  • After quality control, the retained sample contained 81 languages; the analysis used 196 concepts with valid vectors in every retained language.Eleven concepts failed the common-support criterion, leaving 3,240 unordered language dyads; undefined RSA values remained missing.Source: S1 Sample construction and quality… · p. 14

S2 Representational similarity analysis

The analysis compares cross-language semantic relations using RSA, alongside author-defined and reference-style concept groupings. An audit and sensitivity checks support the broad domain contrast but do not establish a unique four-domain taxonomy.

S2 Representational similarity analysis

S3 Primary and reference-style taxonomies

  • The primary taxonomy uses four author-defined operational groups, while a documented reference-style taxonomy draws on Bowern, WOLD, IDS, and WordNet sources.The mappings agree on 185 of 207 input concepts, and the alternative mapping reassigns 22 concepts.Source: S3 Primary and reference-style… · pp. 14–15

S4 Literature motivation for the four domains

  • The source/ontology audit found 89.4% exact agreement, Cohen’s κ = .856, and macro-F1 = .885; endpoint checks supported the broad Nature & Motion versus Socio-Cultural direction, not a unique taxonomy.The complete-case rule determines which items enter the comparable RSA support.Source: S4 Literature motivation for the four… · p. 15, Table S1 · p. 15

S5 Geographic distance and genealogical relatedness

The comparative analyses model semantic dissimilarity across language pairs while accounting for repeated language contributions. Sensitivity analyses show that the cross-model endpoint difference depends on taxonomy.

S5 Geographic distance and genealogical relatedness

  • Genealogical distance is an ASJP lexical–phonological proxy for relatedness, not a formal phylogenetic-tree distance.The measure shares partial material overlap with the semantic analysis because both draw on core-vocabulary traditions.Source: S5 Geographic distance and… · p. 15

S6 Comparative models

  • The pooled model adds encoder, domain, geography, ASJP, and their interactions while retaining the repeated-language structure.Its interaction posteriors, rather than visual comparisons of separate credible intervals, formally test cross-model differences.Source: S6 Comparative models · pp. 15–16
  • Increasing the coefficient-prior scale from 0.2 to 0.5 left the inferential conclusion unchanged: the cross-model endpoint difference remained taxonomy-sensitive.The derived endpoint contrast was 0.0044 [−0.0005, 0.0091] under the primary taxonomy and −0.0029 [−0.0071, 0.0014] under the reference-style taxonomy.Source: S6 Comparative models · pp. 16–17
  • The cross-model difference in the Nature & Motion–Socio-Cultural contrast was not taxonomy-invariant under either prior-scale sensitivity or paired-dyad resampling.The paired bootstrap estimate was 0.0044 (95% interval [0.0012, 0.0074]) under the primary taxonomy and −0.0029 (95% interval [−0.0062, 0.0002]) under the reference-style taxonomy.Source: S6 Comparative models · pp. 16–17, Table S3 · p. 17

S7 Cross-model and representation sensitivity

Across representation and taxonomy checks, the endpoint contrast between Socio-Cultural and Nature & Motion persists, while middle-domain ordering depends on the encoder. Permutation calibration and a consensus subset further qualify the results and their scope.

S7 Cross-model and representation sensitivity

  • Within mBERT, Socio-Cultural < Grammar & Function < Biological Core < Nature & Motion, and the evidence supports a broad positive ASJP association rather than a uniquely grammatical-functional one.The ordering held across full observed inventories; XLM-R-large did not reproduce a consistently positive grammatical increment.Source: S7 Cross-model and representation… · p. 17
  • The middle-domain ordering is checkpoint-dependent, although leave-one-language-out runs preserve each model’s ordering and the middle-domain swap is not driven by one language.Dyad-level agreement is moderate overall and strongest for Socio-Cultural concepts.Source: S7 Cross-model and representation… · p. 17
  • The XLM-R coverage check retained 81 languages, 196 concepts, and all 3,240 dyads, but reused the same support, vectors, and matrices rather than providing independent-data replication.The check documents coverage and provenance.Source: S7 Cross-model and representation… · p. 17, Table S4 · p. 17

S8 Concept-label permutation calibration

  • The concept-label permutation calibration found observed mean RSA dissimilarity below every replicate across all 16 encoder–taxonomy–domain combinations.With the add-one correction, the lower-tail probability was .0005 and the two-sided probability was .0010; this calibration does not test causal interpretations or make dyad-level entries independent.Source: S8 Concept-label permutation calibration · pp. 17–18

S9 Consensus-subset sensitivity

  • On the 175-concept consensus subset, Socio-Cultural dissimilarity remained below Nature & Motion in both encoders, but the complete middle-domain ordering was not invariant.The contrast was .888 versus .935 for mBERT and .938 versus .977 for XLM-R-large.Source: S9 Consensus-subset sensitivity · p. 18

S10 Dependence and domain-size sensitivity

Robustness checks support the Socio-Cultural–Nature & Motion endpoint contrast, while the complete domain ordering and some geographic associations vary with analytical choices. The checks address domain profiles and cross-model comparisons under alternative classifications, sampling procedures, and model specifications.

S10 Dependence and domain-size sensitivity

  • The Socio-Cultural–Nature & Motion ordering held in 89.6% and 95.8% of equal-size subsamples, while the full four-domain ordering held in 33.6% and 35.4%.These are archived mBERT sensitivity checks; the endpoint contrast was more stable than the complete ordering.Source: S10 Dependence and domain-size… · p. 18, Table S5 · p. 19

Reference-style classification

  • Under the reference-style mapping, Nature & Motion means exceeded Socio-Cultural means for mBERT (.921 vs .883) and XLM-R-large (.973 vs .938), while middle-domain order changed.The mapping reassigns 22 of 207 input concepts; the cross-model difference in endpoint contrasts was −.0029 (95% CrI [−.0071, .0014]).Source: Reference-style classification · p. 18

Leave-one-language-out summaries

  • Each model retained its full-inventory domain ordering in all 81 leave-one-language-out summaries, and no single language drove the endpoint contrast.The omitted-language means are displayed as error bars in the main-text cross-model domain-profile figure.Source: Leave-one-language-out summaries · p. 18

Paired dyad bootstrap

  • Paired-dyad bootstrap estimates differed by taxonomy, while full-support pooled posterior intervals included zero under both taxonomies; the bootstrap does not establish an invariant cross-model difference.Bootstrap contrasts were .0044 (95% interval [.0012, .0074]) for the primary taxonomy and −.0029 ([-.0062, .0002]) for the reference-style taxonomy.Source: Paired dyad bootstrap · p. 18

Frequency-adjusted secondary models

  • Frequency-adjusted item-level models used English Zipf frequency, but their differing outcome and estimands do not establish that the RSA domain contrast is unchanged after adjustment.The frequency coefficient was negative for mBERT (−.02620, 95% Wald interval [−.03228, −.02012]) and XLM-R-large (−.01414, [−.02424, −.00404]).Source: Frequency-adjusted secondary models · p. 19

Layer diagnostics and equal-size subsampling

  • Archived layer diagnostics cover Grammar & Function and Socio-Cultural, so they do not establish whether the endpoint contrast persists across layers.These checks are not matched fresh-vector, dual-encoder validation.Source: Layer diagnostics and equal-size… · p. 19

Concept-level profiles and triplet diagnostics

Summary

  • Across robustness checks, the Socio-Cultural–Nature & Motion endpoint contrast and positive ASJP association persist across both models, but the complete ordering and some geographic associations remain choice-sensitive.The endpoint contrast is comparatively stable; the exact four-domain ordering is more conditional on domain assignment, model, and included concepts.Source: Summary · p. 19

S13 Reproducibility materials

The reproducibility package documents data, analyses, diagnostics, and software provenance, but does not include checkpoint weights or complete weight-hash coverage.

S13 Reproducibility materials

  • The package includes concept mappings, quality-control records, model outputs, sensitivity analyses, diagnostics, source code, seeds, and hashes, but not checkpoint weights.The XLM-R-large manifest marks its weight file as not hashed, so complete weight-hash coverage is not claimed.Source: S13 Reproducibility materials · p. 19

S14 Supplementary Tables

The supplementary tables document concept classification, model estimates, sensitivity analyses, and analysis provenance. They include complete-support results and descriptive layer-wise diagnostics.

S14 Supplementary Tables

  • Tables S9–S11 document domain concept counts, concepts reassigned by the reference-style mapping, and dyad-level agreement between taxonomies.The counts precede complete-case exclusion, and taxonomy-agreement bootstrap intervals use 2,000 concept-level resamples.Source: Table S9 · p. 20, Table S10 · p. 20, Table S8 · p. 20
  • Tables S12–S17 report global and fixed-effect models, adjusted means, geography and genealogical-distance estimates, and sampling diagnostics.The adjusted means are reconstructed from frozen full-support fits; pooled rows use the complete 25,920-row support.Source: Table S12 · p. 21, Table S13 · p. 21, Table S14 · p. 22
  • Under the reference-style XLM-R analysis, adjusted means were .973 for Nature & Motion and .938 for Socio-Cultural; pooled endpoint intervals included zero under both taxonomies.The primary pooled contrast was 0.0044 [−0.0003, 0.0090], and the reference-style contrast was −0.0029 [−0.0071, 0.0014].Source: S14 Supplementary Tables · p. 23
  • Table S18 reports descriptive layer-wise slopes for Grammar & Function and Socio-Cultural; Table S19 reports frequency-adjusted item-level models for both encoders.The layer-wise intervals are not additional confirmatory tests; the frequency models use 635,040 observations per encoder and 95% Wald intervals.Source: Table S18 · p. 24, Table S19 · p. 24
  • Tables S20–S25 document retained languages, XLM-R extraction provenance, fixed-sample means and leave-one-out ranges, encoder agreement, stability, and software provenance.The fixed-sample and leave-one-language-out summaries use the same 81 languages, 196 concepts, and 3,240 dyads.Source: Table S20 · p. 24, Table S21 · p. 26, Table S22 · p. 26
  • Table S26 provides the complete study-concept mapping, while Table S27 lists pooled two-domain contrasts under default and sensitivity priors.The mapping distinguishes the study-specific alignment from the original concept lists and retains a separate frozen macro-domain rule.Source: Table S26 · p. 28, Table S27 · p. 31

S15 Encoder comparison details

The encoder comparison uses a shared analytic support and workflow, while archived-vector and frequency-adjusted analyses serve as secondary sensitivity checks.

S15 Encoder comparison details

  • Both encoders use the same languages, concepts, mappings, lexical input rule, pooling, centering, RSA construction, and predictor matrices.Domain fits contain 12,960 rows each and the pooled model contains 25,920 rows.Source: S15 Encoder comparison details · p. 32
  • Frequency-adjusted models use 635,040 item-level concept–dyad observations per encoder, English Zipf frequency, and exact Gaussian REML with marginalized language and concept effects.These models use item-level cosine distances.Source: S15 Encoder comparison details · p. 32
  • Equal-size subsampling and the 13-layer table use archived mBERT vectors and are preliminary sensitivity analyses.The source taxonomy review forms document operational classifications.Source: S15 Encoder comparison details · p. 32

S16 Supplementary figures

Figure S1 presents mBERT concept-level heterogeneity through item-level coefficients and a full-inventory representational profile.

  • Panel (a) reports item-level regression coefficients for individual concepts.Source: Figure S1 · p. 32
  • Panel (b) summarizes the full-inventory mBERT representational profile, with arrows indicating observed associations.Source: Figure S1 · p. 32
  • The geographic contrast shown is specific to the mBERT primary-taxonomy configuration.Source: Figure S1 · p. 32

Related papers

Loading 2610.30778v1…