On the universal structure of human lexical semantics

Significance

Semantics, or meaning expressed through language, provides indirect access to an underlying level of conceptual structure. To what degree this conceptual structure is universal or is due to properties of cultural histories, or to the environment inhabited by a speech community, is still controversial. Meaning is notoriously difficult to measure, let alone parameterize, for quantitative comparative studies. Using cross-linguistic dictionaries across languages carefully selected as an unbiased sample reflecting the diversity of human languages, we provide an empirical measure of semantic relatedness between concepts. Our analysis uncovers a universal structure underlying the sampled vocabulary across language groups independent of their phylogenetic relations, their speakers’ culture, and geographic environment.

Abstract

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 indirect 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 to translate words to and from languages carefully selected to be representative of worldwide diversity. These translations reveal cases where a particular language uses a single “polysemous” word to express multiple concepts that another language represents using distinct words. We use the frequency of such polysemies linking two concepts as a measure of their semantic proximity and represent the pattern of these linkages by a weighted network. This network is highly structured: Certain concepts are far more prone to polysemy than others, and naturally interpretable clusters of closely related concepts emerge. Statistical analysis of the polysemies observed in a subset of the basic vocabulary shows that these structural properties are consistent across different language groups, and largely independent of geography, environment, and the presence or absence of a literary tradition. The methods developed here can be applied to any semantic domain to reveal the extent to which its conceptual structure is, similarly, a universal attribute of human cognition and language use.

Constructing Semantic Network from Translations

We represent semantic relations obtained from dictionary translations of the chosen concepts as a network. Two meanings are linked if they can be reached from one another by a translation into some language and then back. The link is weighted by the number of such paths (of length 2), that is, the number of polysemous words that represent both meanings (details are provided in Materials and Methods). Fig. 1 illustrates the construction with examples from two languages. Translating the word SUN into Lakhota results in wí and áηpawí. Although the latter picks up no other meaning, wí is polysemous: it possesses additional meanings of MOON and month, so they are linked to SUN in the network. A similar polysemy is observed in Coast Tsimshian, where gyemk, the translation of SUN, also means heat, thus providing a link between SUN and heat. We write the initial Swadesh concepts (SUN and MOON in this example) in capital letters, whereas other concepts that arise through translations (month and heat here) are written in lowercase letters. We restrict our study to the neighborhood of the initial Swadesh concepts, so further translations of these latter concepts are not followed.

With this approach, we can construct a semantic network for each individual language. It is conceivable, however, that a group of languages bears structural resemblances as a result of the speakers’ sharing common historical or environmental features. A link between SUN and MOON, for example, recurs in both languages illustrated in Fig. 1, yet does not appear in many other languages, where other links are seen instead. Thus, for example, SUN is linked to divinity and time in Japanese and to thirst and DAY/DAYTIME in the Khoisan language !Xóõ. The question is then the degree to which these semantic networks are similar across language groups, reflecting universal conceptual structure, and the extent to which they are sensitive to cultural or environmental variables, such as phylogenetic history, climate, geography, or the presence of a literary tradition. We test such questions by constructing aggregate networks from groups of languages that share a common cultural and environmental property and comparing these networks between different language groups.

Semantic Clusterings

As a point of comparison for the networks obtained from such groups of languages, we show the network obtained from the entire set of languages in Fig. 2 and SI Appendix, Fig. S6, displaying only the links that appear more than once. This network exhibits the broad topological structure of polysemies observed in our data. It reveals three almost disconnected clusters of concepts that are far more prone to polysemy within each cluster than between them. These clusters admit a natural semantic interpretation. Thus, for example, the semantically most uniform cluster, colored in blue, includes concepts related to water. A second, smaller cluster, colored in yellow, groups concepts related to solid natural materials (e.g., STONE/ROCK, MOUNTAIN) and associated landscape features (e.g., forest, clearing, highlands). The third cluster, colored in red, is more diverse, containing terrestrial terms (e.g., field, floor, ground, EARTH/SOIL), celestial objects [e.g., CLOUD(S), SKY, SUN, MOON], and units of time (e.g., DAY, NIGHT, YEAR). Although the clustering is strong, there do exist rare polysemies that occur only once in our dataset (and are thus not displayed in Fig. 2) connecting the three clusters. Thus, for example, CLOUD(S) is polysemous with lightning in Albanian, whereas the latter is polysemous with STONE/ROCK in !Xóõ, and whereas holy place is a polysemy for MOUNTAIN in Kisi, it is instead polysemous with LAKE in Wintu. The individual networks including such weak links can be accessed in our web-based platform (29).

The links defining each of the three clusters can be understood in terms of well-known kinds of polysemies: metonymies (polysemy between part and whole) and commonly found semantic extension to hypernyms (more general concepts), hyponyms (more specific concepts), and cohyponyms (specific concepts belonging to the same category). The first cluster contains both liquid substances and topographic features metonymically related to water. The substance polysemies in this cluster are various liquids, cohyponyms of WATER. The topographic polysemies (e.g., LAKE, RIVER) are also linked as cohyponyms under “body of water” and “flowing water.” Similarly, in the third cluster, the bridge between the terrestrial and celestial components is provided by the hyponyms of “granular aggregates,” which span both the terrestrial EARTH/SOIL, DUST, and SAND and the airborne SMOKE and CLOUD(S).

Evidence for Universal Semantic Structure

The semantic network across languages reveals a universal set of relationships among these concepts that possibly reflects human conceptualization of these semantic domains (8, 12, 30). Alternatively, it has been postulated that such semantic relations are strongly influenced by the physical environment that human societies inhabit (31).

To address this question, we group the languages by various factors (SI Appendix, Table SIII) comprising the geography, climate, or topography of the region where they are spoken, and the presence or absence of a literary tradition in them, and we test the effect of these factors on the semantic network. We measured the similarity between these groups’ semantic networks in several ways. First, we measured the correlation between the commute distances (32) between nearby concepts (Materials and Methods and SI Appendix, section III A 1). Then, to compare these networks’ large-scale structure, we clustered the concepts in each network hierarchically as a dendrogram (SI Appendix, Fig. S8) and compared them using two standard tree metrics (33–35): the triplet distance (Dtriplet) and the Robinson–Foulds distance (DRF) (Materials and Methods and SI Appendix, section III A). To test whether these networks are more similar than what we would expect by chance, we performed bootstrap experiments, where we compared each network with the one where the concepts were randomly permuted (SI Appendix, section III). As shown by the p1 values in Fig. 3, in every case, the networks of real language groups are far more similar to each other than to these randomly permuted networks, allowing us to reject decisively the null hypothesis that these semantic networks are completely uncorrelated (statistical details are provided in SI Appendix, section III B).

All these tests thus establish that different language groups do indeed have semantic structure in common. To explore this universal semantic structure further, we tested a null hypothesis at the other extreme, that cultural and environmental variables have no effect on the semantic network. For this purpose, we performed a different kind of bootstrap experiment, where we replaced each language group with a random sample of the same size from the set of languages. As denoted by p2 in Fig. 3, we find that, with rare exceptions, there is no statistical support (SI Appendix, section III B) for the hypothesis that the differences between the language groups studied are any larger than between random groups of the same size. This fact means that the impacts of cultural and environmental factors are weaker than what can be established with our dataset; thus, our results are consistent with the hypothesis that semantic clustering structure is independent of culture and environment in these semantic domains.

Heterogeneity of the Semantic Network

The universal semantic network shown in Fig. 2 is heterogeneous in both node degrees and link weights. The numbers of polysemies involving individual meanings are uneven, possibly trending toward a heavy-tailed distribution (Fig. 4). This distribution indicates that concepts have different tendencies of being polysemous. For example, EARTH/SOIL has more than 100 polysemies, whereas SALT has only a few.

Interestingly, we find that this heterogeneity is also universal: The numbers of polysemies of the various concepts that we studied in any two languages are strongly correlated with each other. This correlation holds despite the observation that the languages differ in the overall magnitude of polysemy, so that the same concepts are far more polysemous in some languages than in others (SI Appendix, Fig. S2). In fact, a simple formula predicts the number of polysemies, nSL, involving sense S in language L rather well (SI Appendix, Fig. S9):

where nS is the number of polysemies involving sense S in the aggregate network from all languages, N is the total number of polysemies in this aggregate network, and nL is the number of polysemies in the language L. This formula is exactly what we would expect (Materials and Methods and SI Appendix, section II B) if each language randomly and independently draws a subset of polysemies for each concept S from the universal aggregate network, which we can identify as an underlying “universal semantic space” (SI Appendix, Fig. S5). The data for only three concepts, MOON, SUN, and ASHES, deviate from this linear pattern by more than the expected sampling errors (p≈0.01) in that they display an initial rapid increase in nSL with nL, followed by a saturation or slower increase at larger values of nL (SI Appendix, Fig. S10). These deviations can be accommodated using a slightly more complicated model described in SI Appendix, section IV.

Discussion

We propose a principled method to construct semantic networks linking concepts via polysemous words identified by cross-linguistic dictionaries. Based on the method, we found overwhelming evidence that the semantic networks for different groups share a large amount of structure in common across geographic and cultural differences. Indeed, our results are consistent with the hypothesis that cultural and environmental factors have little statistically significant effect on the semantic network of the subset of basic concepts studied here. To a large extent, the semantic network appears to be a human universal: For instance, SEA/OCEAN and SALT are more closely related to each other than either is to SUN, and this pattern is true for both coastal and inland languages.

These findings have broad implications. Universal structures in lexical semantics such as we observe can greatly aid reconstruction of human history using linguistic data (37, 38). Much progress has been made in reconstructing the phylogenies of word forms from known cognates in various languages, thanks to the ability to measure phonetic similarity and our knowledge of the processes of sound change. The relationship between semantic similarity and semantic shift, however, is still poorly understood. The standard view in historical linguistics is that any meaning can change to any other meaning (39, 40), and no constraint is imposed on what meanings can be compared with detect cognates (41). In contrast to this view, we find that at least some similarities occur in a heterogeneous and clustered fashion.

Previous studies (9, 19–21, 23, 24, 42–45) have investigated the presence or absence of universality in how languages structure the lexicon in a few semantic domains dealing with personal items like body parts, perceptual elements like color metaphors, and cultural items like kinship relations. In this work, we study instead the domain of celestial and landscape objects that one may a priori expect to be strongly affected by the environment. We find, however, that the semantic networks on which these natural objects lie are universal. It is generally accepted among historical linguists that language change is gradual: Over historical time, words gain meanings when their use is extended to similar meanings and lose meanings when another word is extended to the first word’s meaning. If such transitional situations are common among polysemies, then the meaning shifts in this domain are likely to be equally universal, and the observed weights on different links of the semantic network reflect the probabilities of words to be in transition. As such, semantic shifts can be modeled as diffusion in the conceptual space, or along a universal semantic network, and thus our constructed networks can serve as important input to methods of inferring cognates.

Our results are obtained from detailed typological studies from a sample of the world’s languages. We chose to collect data manually from printed dictionaries. This approach ensures that our sample is unbiased and representative of the variation known among languages but foregoes the large sample size that online digital resources offer, because these data are dominated by a few languages from developed countries with long-established writing systems and large speaker populations. We find, however, that the patterns of polysemy in our data have little correlation with environmental, social, and other linguistic attributes of the language. This consistency across language groups suggests that languages for which digital resources are available are likely to produce networks similar to those networks created from the full sample. Therefore, the semantic network constructed here can be extended with more extensive data from online dictionaries and digital corpora by automated means (46). In such analysis with digitally available resources, one can examine if patterns of polysemy could be shared among more closely related language groups than the genus level, and if universality holds for other semantic domains.

Materials and Methods

Supplementary Material

Acknowledgments

We thank Ilia Peiros, George Starostin, Petter Holme, and Laura Fortunato for helpful comments. H.Y. acknowledges support from Complex Agent Based Dynamic Networks (CABDyN) Complexity Center and the support of research grants from the National Science Foundation (Grant SMA-1312294). W.C. and L.S. acknowledge support from the University of New Mexico Resource Allocation Committee. T.B., J.F.W., E.S., C.M., and H.Y. acknowledge the Santa Fe Institute and the Evolution of Human Languages program. The project and C.M. are also supported, in part, by the John Templeton Foundation.

Footnotes

References

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.