PaperlayerView exact revision on arXiv

Source-linked AI summary

On the universal structure of human lexical semantics

Hyejin Youn, Logan Sutton, Eric Smith, Cristopher Moore, Jon F. Wilkins, Ian Maddieson, William Croft, Tanmoy Bhattacharya

arXiv:1504.07843v1physics.soc-phcs.CL

TL;DR

The paper asks whether conceptual relationships are general or vary with environmental, historical, or linguistic context. It measures concept proximity using cross-language patterns of polysemy and represents those relations in a network. The authors report universal intrinsic closeness between concepts, alongside uneven distribution of salient concepts across conceptual space.

  • Problem

    The authors ask whether observed conceptual relationships are general or sensitive to environment, language phylogeny, and intrinsic linguistic factors.

    Source: RESULTS · p. 4

  • Method

    The authors use weighted node degree as a proxy for a concept’s polysemy tendency and total link weight as a proxy for a language’s tendency to distinguish word senses.

    Source: RESULTS · p. 7

  • Results

    Polysemy-based similarity relations exhibit universal intrinsic closeness, while salient concepts are unevenly distributed across conceptual space.

    Source: DISCUSSION · p. 7

  • Takeaways & Limitations

    The findings support treating concept closeness in the studied vocabulary as part of a universal structure.

    Source: DISCUSSION · p. 7

    Comparisons between these languages and the other languages in the sample could be problematic.

    Source: Polysemy data · p. 9

Abstract

from arXiv · show

How universal is human conceptual structure? The way concepts are organized in the human brain may reflect distinct features of cultural, historical, and environmental background in addition to properties universal to human cognition. Semantics, or meaning expressed through language, provides direct access to the underlying conceptual structure, but meaning is notoriously difficult to measure, let alone parameterize. Here we provide an empirical measure of semantic proximity between concepts using cross-linguistic dictionaries. Across languages carefully selected from a phylogenetically and geographically stratified sample of genera, translations of words reveal cases where a particular language uses a single polysemous word to express concepts represented by distinct words in another. We use the frequency of polysemies linking two concepts as a measure of their semantic proximity, and represent the pattern of such linkages by a weighted network. This network is highly uneven and fragmented: certain concepts are far more prone to polysemy than others, and there emerge naturally interpretable clusters loosely connected to each other. Statistical analysis shows such structural properties are consistent across different language groups, largely independent of geography, environment, and literacy. It is therefore possible to conclude the conceptual structure connecting basic vocabulary studied is primarily due to universal features of human cognition and language use.

RESULTS

The language-group comparisons support shared semantic structure, while the pooled network reveals interpretable clusters and uneven polysemy across concepts. A random-sampling model captures much of this variation, with notable exceptions.

RESULTS

  • SUN links to MOON in some languages, but elsewhere to divinity and time in Japanese or thirst and DAY/DAYTIME in !X´o˜o.The SUN–MOON link recurs in two languages but is absent in many others.Source: RESULTS · p. 4
  • The network has three almost-disconnected clusters: water, solid natural substances around STONE/ROCK and MOUNTAIN, and a looser cluster bridging terrestrial and celestial concepts.The figure marks the clusters in blue, yellow, and red, respectively.Source: RESULTS · p. 4, Figure 2 · p. 5
  • Language-group dendrograms were closer to one another than to randomly permuted trees, while their distances were statistically indistinguishable from bootstrap resamples.The analyses conclude that languages share semantic structure without significant environmental or cultural influence.Source: RESULTS · pp. 5–6, Figure 3 · p. 6
  • Concepts differ sharply in polysemy: EARTH/SOIL has hundreds of polysemes, whereas SALT has only a few.The counts are uneven and may follow a heavy-tailed distribution.Source: RESULTS · p. 6
  • Weighted node degree proxies a concept’s polysemy tendency, while total link weight proxies a language’s tendency to distinguish word senses.The random-sampling model captures the data’s gross features, but MOON, SUN, and ASHES show nonlinear increases; saturation parameters yield good agreement while keeping the two parameters independent.Source: RESULTS · p. 7

DISCUSSION

The paper treats cross-linguistic polysemy links as evidence of intrinsic conceptual closeness that transcends cultural and environmental differences. It also connects this structure to uneven concept density and possible applications to semantic change and cognate inference.

DISCUSSION

  • Cross-linguistic polysemy links exhibit a universal structure of intrinsic conceptual closeness that transcends cultural and environmental differences.Source: DISCUSSION · p. 7
  • Highly variable polysemy rates indicate concepts are unevenly distributed in conceptual space, with a concept’s propensity for polysemy interpreted as local density.The model suggests that, given overall semantic ambiguity in languages, local density determines the degree of polysemy.Source: DISCUSSION · pp. 7–8
  • If meaning change is gradual, link weights reflect how often transitions between meanings are captured, and semantic shifts can be modeled as diffusion through conceptual space.Source: DISCUSSION · p. 8
  • The constructed polysemy networks could serve as input to methods for inferring cognates.Source: DISCUSSION · p. 8
  • The absence of significant cladistic correlation suggests archived dictionaries could extend the conceptual space without strong language-family bias.The paper also identifies automated construction of semantic spaces in unexplored domains as a possible use of large-corpus samples.Source: DISCUSSION · p. 8

Polysemy data

The study identifies polysemies by comparing translations of basic vocabulary across a stratified sample of languages. It uses modern European languages as a semantic metalanguage and reports that lexicographic practice addressed a potential ambiguity in those translations.

Polysemy data

  • The object languages were selected from a phylogenetically and geographically stratified sample of low-level language families or genera, and translations were obtained for 22 Swadesh basic-vocabulary word senses.Source: Polysemy data · p. 8
  • Researchers retained translation synonyms and identified polysemies by back-translating object-language terms through modern European-language dictionaries.Source: Polysemy data · pp. 8–9
  • In Figure 4, strength counts polysemies while degree counts linked concepts; heaven has the most polysemies, mostly with SUN, but degree three.Source: Figure 4 · p. 9
  • Polysemy in the semantic-metalanguage languages could complicate comparisons, but lexicographers annotated such cases and standard practice was reported sufficient.Source: Polysemy data · p. 9

Comparing semantic networks between language groups

The authors compare semantic-network dendrograms across language subgroups using tree distances and bootstrap tests. Most subgroup pairs show more similarity than the null trials, while comparisons based on nonlinguistic groupings are not statistically distinguishable from random sampling in this dataset.

Comparing semantic networks between language groups

Null model

The null model treats concept–language polysemy counts as independent and estimates their expected frequencies from concept- and language-level tendencies. The study tests this model against observed data and selects 22 basic lexical concepts from natural and geographic domains for the network analysis.

Null model

  • The model treats concepts as independent and estimates each concept–language polysemy count from the concept’s and language’s respective tendencies to participate in polysemy.Those tendencies are estimated from the observed marginal fractions of polysemy associated with each concept and language.Source: Null model · pp. 10–11
  • The model is tested by comparing observed and simulated Kullback–Leibler divergence values, with the p-value defined by the simulated probability of a value at least as large as observed.Simulated ensembles are generated according to the model’s probabilities.Source: Null model · p. 11

A. Criteria for selection of meanings

  • The translations use lexical concepts rather than grammatical inflections or function words, and body parts are omitted because they have been studied previously.Body parts otherwise meet the stated selection criteria for a connected domain.Source: A. Criteria for selection of meanings · p. 14
  • The analysis uses 22 lexical concepts from basic vocabulary, selected for likely single-word representation and concrete referents in natural or geographic settings.The selected domains are intended to yield a connected graph rather than many disconnected singletons.Source: A. Criteria for selection of meanings · p. 14
  • The selected concepts cover celestial phenomena and time units, natural features, and material substances.Examples include STAR, SUN, MOON, YEAR, SKY, RIVER, STONE/ROCK, FIRE, and WATER.Source: A. Criteria for selection of meanings · p. 14

B. Language List

The language list is organized by language groups, with genus-level classification emphasized as generally agreed upon. The accompanying table also describes how language subsets based on nonlinguistic variables are compared.

B. Language List

  • The notes caution that Papuan languages are not generally treated by historical linguists as one phylogenetic group, while other listed families vary in acceptance.Oceania in the table includes Southeast Asia.Source: B. Language List · p. 14
  • The study’s language list is presented in Table I, whose caption notes that genus-level classification is generally agreed upon.The accompanying notes distinguish the status of higher-level family classifications.Source: B. Language List · p. 14, Table I · p. 15
  • Table II groups languages by nonlinguistic variables and defines network differences as tree distances between dendrograms of Swadesh words generated by spectral clustering.Source: Table II · p. 16

C. Language groups

The authors test whether the polysemy-network structure depends significantly on typological features. The tested features span literacy, geography, topography, and climate.

C. Language groups

  • The authors test whether polysemy-network structure depends in a statistically significant way on typological features.Source: C. Language groups · p. 16
  • The tested features include the presence or absence of a literary tradition, geography, topography, and climate.Source: C. Language groups · p. 16
  • The feature list reports the number of languages represented for each typological feature.Source: C. Language groups · p. 16

1. Aggregation of language samples

The authors assess whether language samples can support an aggregate network by testing an explicit random-sampling model rather than relying only on summary statistics. They report that the dataset is typical by their measures, supporting use and interpretation of the aggregate graph.

1. Aggregation of language samples

  • The authors use an explicit algorithm to assign probabilities to samples of Swadesh node degrees because summary statistics imply generating processes that must be tested.Source: §1 · p. 17
  • The dataset is typical by the reported significance measures, supporting use and interpretation of the aggregate graph (Fig. 2).The measures are designed to be consistent with both the aggregate graph and the sampling algorithm.Source: §1 · p. 17
  • The sampling model calibrates a general mean-degree distribution to reproduce sample means based on both Swadesh-entry and language-weight properties.Source: §1 · p. 17
  • The random-sampling consistency criterion depends on scale: uncorrelated synonymous polysemies may make individual languages nonrandom while small language groups converge toward random-sample consistency.The authors therefore identify the smallest groupings consistent with randomness and model excess but uncorrelated synonymous polysemy within individual languages.Source: §1 · p. 17

2. Independent sampling from the aggregate graph

This section evaluates whether aggregate link patterns can be treated as independent random samples by comparing observed configurations with candidate sampling models using KL divergence. It examines how nine- and three-language binning and language size relate to the patterns’ typicality under saturation sampling.

Product model with intrinsic property of concepts

  • The interaction model accepts the 9-language aggregate and, after a small language-size adjustment, the 3-language aggregate, but strongly excludes the joint configuration of all 81 languages.Source: §2 · p. 20

Product model with saturation

  • The saturation model addresses product-model overprediction by bounding each concept’s polysemy count at a saturation level interpreted as its number of proximal concepts.Source: §2 · pp. 20–21
  • Saturation varies across entries: WATER shows no significant deviation from the product model, whereas MOON has the lowest saturation value, approximately 3.4.Source: Figure 6 · p. 21
  • Fit parameters were generally insensitive to bin size, but MOON’s saturation estimate was unstable in single-language bins because one Itelman language with one link had the likely moon/month polysemy.Source: §2 · p. 22
  • At 1%, the product model is rejected for 9-language bins, while the saturation model is typical of the same configuration at approximately 59%.Source: Figure 7 · p. 23

2. Independent sampling from the aggregate graph

  • Individually solved correction factors ϕL vary smoothly with language size nL, potentially refining the product model or reflecting small-language degree-distribution interactions with the error function.Source: §2 · p. 23, Figure 8 · p. 24
  • Nine-language binning makes the data typical under saturation sampling (P ∼59%), while the product model is excluded below 1%; three-language bins remain plausible (P ∼7%), whereas product sampling is rejected at p ≈10^-3.The product model preserves the marginals; the finer test uses 27 three-language bins.Source: §2 · pp. 24–25, Figure 9 · p. 24
  • The three-language-bin acceptance suggests language size nL is an excellent explanatory variable, with residual language variations canceling approximately even in small aggregations.Source: §2 · p. 25

3. Single instances as to aggregate representation

This section examines how polysemy-link variation differs between individual languages and language aggregates, and whether clumpy link sampling accounts for the pattern. It compares observed regression behavior with expectations from the fitted saturation model and clumped probabilities in three-language bins.

Power tests and uneven distribution of single-language p-values

  • The saturation model is weakly plausible for bins as small as three languages, and these analyses do not require deviations from independent link sampling at that aggregation scale.The preceding aggregate-scale analysis found no need for departures from independent sampling down to this bin size.Source: §3 · p. 25

Correlated link assignments

  • Models of non-independent, clumpy link sampling reproduce single-language deviations while preserving the aggregate mean pattern.The section also interprets the disappearance of clumping after binning as evidence that clumps are uncorrelated among languages.Source: §3 · pp. 25, 28, 31

Power tests and uneven distribution of single-language p-values

  • Approximately 4–5 of 81 languages exceed the saturation-model expectation at p-values 0.05 and 0.1, while another 2–3 are rejected by the product model near p = 0.2.The 27 three-language aggregate distributions are marginally consistent with random sampling under the saturation model.Source: §3 · pp. 25–26, Figure 11 · p. 27

Excess fluctuations in degree of polysemy

  • Three-language bins are consistent with Poisson sampling across link counts, whereas single-language relative variance rises with language size, consistent with clumpy link assignment.The disappearance of the variance trend after binning indicates that the clumps are uncorrelated across languages.Source: §3 · pp. 27–28, Figure 13 · p. 29

Single instances as to aggregate representation

  • The clumped-probability samples retained significant probability for three-language-binned regressions around zero without ex-post fitting.The privileged-word set was drawn independently for each language.Source: §3 · pp. 31–32
  • Simulations from model (15) produced three-language-bin regression coefficients consistent with zero in about one-third of cases.Typically, the three-bin coefficient decreased by about one-third from the single-language regression.Source: §3 · p. 31
  • The observed regression coefficient of 0.012 lies far outside multinomial-sampling support for the fitted psat S|L model, while clumped probabilities produce three-language-binned regressions around zero.Source: §3 · pp. 31–32
Loading 1504.07843v1…