The typology of sound symbolism: Defining macro-concepts via their semantic and phonetic features

3.1 Establishing near-universal vocabulary

When searching for sound symbolic patterns, basic vocabulary is especially suitable since it consists of concepts that are supposed to be salient for all speakers regardless of the language, culture and era. These concepts broadly relate to the fundamental categories of the mind (e. g. emotions, senses, tastes, perceivable physical properties), the body (e. g. body part terms, mental and bodily functions), society (e. g. kinship terms, human categories), the surrounding world (e. g. natural entities) and reference (e. g. deictic concepts, determiners, spatial relations). Furthermore, to account for language-specific delimitation of semantic fields, boundaries between concepts were generalized according to prevailing typological and physiological patterns. For example, singular-plural distinctions were included for pronouns but not dual, paucal etc., and individual terms for ‘hand’ and ‘arm’ were included rather than a term for the entire limb. Thus, the selection of the 344 featured concepts (see Table 2) was based on:

1. near-universality, i. e. presence in the majority of the world’s languages,

2. strong linguistic typological patterns as a basis for drawing the borders between concepts,

3. physiological and natural constraints as a basis for drawing the borders between concepts,

4. lists of basic vocabulary for high comparability with similar studies.

To begin with, we included all of the 56 fundamental oppositional concepts that were shown to have great sound symbolic potential by Johansson (2017), mostly based on proposed lexical universals (Dixon 1982; Goddard 2001; Goddard & Wierzbicka 2002; Koptjevskaja-Tamm 2008; Paradis et al. 2009), namely i-you, big-small, good-bad, this-that, many-few, before-after, above-below, far-near, man-woman, black-white, hot-cold, here-there, long-short, night-day, full-empty, new-old, round-flat, dry-wet, wide-narrow, thick-thin, smooth-rough, heavy-light, dark-light, quick-slow, hard-soft, deep-shallow, high-low and mother-father.

Associations between colors and sounds are perhaps among the most commonly studied sound symbolic areas. Even though synesthetes experience these associations more strongly than non-synesthetes (Ward et al. 2006), hue, chroma and lightness are also associated with auditory frequency and loudness (Spence 2011; Walker 2012; Hamilton-Fletcher et al. 2017) and other senses, such as touch (Ludwig & Simner 2013), in the general population. What is more, Berlin and Kay (1969) famously demonstrated strong cross-linguistic regularities in the lexicalization patterns of monolexemic color terms. However, monolexemic terms are not guaranteed to be free of lexical interference from non-color concepts, as they can often be traced back to old derivations of a referent in nature having that particular color. For example, ‘green’ is ultimately derived from Proto-Indo-European *ǵʰreh₁ -ni-, meaning ‘to grow’ (Kroonen 2010–), i. e. ‘plant-colored’. Selecting color concepts based on these lexicalization patterns may therefore not be ideal for the purposes of this paper. Therefore, selecting these concepts based on color opponency (Kay & Maffi 2013) was judged a more suitable choice since it offers a more neutral division of the color spectrum which is also cross-linguistically grounded. Thus, we included two pairs of fundamental opponent chromatic colors (red-green and yellow-blue), a single pair of fundamental achromatic colors (black-white), and gray, the combination of the most basic colors. Number concepts were also narrowed down in a similar manner to accommodate most of the world’s numeral systems (Comrie 2013), namely decimal (base 10), vigesimal (base 20), restricted (individual terms up to around ‘five’ which are combined to create higher numbers) and extended body-part (based on individual words for body part without an arithmetic base). The final selection included numbers one through twelve, as well as twenty.

Among the deictic concepts, first, second and third personal pronouns in the singular and plural, as well as inclusive and exclusive first-person plural (1SG, 2SG, 3SG, 1PLI, 1PLE, 2PL, 3PL). These general concepts were included to account for the various strategies that languages use to divide pronouns and nouns into noun classes and grammatical genders (Corbett 2013) as the alternative would force us to include separate slots for all possible categories (such as masculine, feminine, neuter, common, animate, inanimate, human, non-human, countable, uncountable etc.) for each pronoun concept.

Proximal, medial and distal location adverbs (here, there, there yonder) and demonstratives in the singular (this, that, that yonder) were chosen for being the most common types cross-linguistically (Diessel 2014), with the addition of two temporal deictic concepts (now, then) and six interrogative pronouns, which incorporate the notions of human (who?), non-human (what?), location (where?), time (when?), manner (how?), and reason (why?).

Closely related to demonstratives, location concepts seem to be arranged to be maximally informative within languages, i. e. languages seem to categorize objects in a way that favors accurate mental reconstruction by a listener of a speaker’s intended meaning rather than basing it on other natural or salient categories (Khetarpal et al. 2013). Despite this, there does not seem to be a clear-cut set of universal categories (Levinson & Meira 2003; Burenhult & Levinson 2008; Khetarpal et al. 2010). Thus, the selected concepts were only meant to convey immediate relational positions to objects rather than directions (e. g. above but not up), and belonged to four types: horizontal (left-right, behind-in front of, beside), vertical (above-below), time (before-after), and object-related (inside-outside, between). In addition, a universal (all), existential (some) and negatory (nothing) quantifier were included, as well as an equal (same) and contrastive (other) determiner.

Linguistic variation in age categories creates similar issues when working with cross-linguistic data, e. g. the Austroasiatic language Khmu [kgj] distinguishes about twice as many categories as English [eng] (children, teenagers, young adults, adults and elders); hence, the selected concepts only included a general term (person) and three age-coded groups of concepts in order to fit most languages: elderly (old man, old woman), adult (man, woman) and child (boy, girl).

Surprisingly, one of the most studied anthropological subjects, kinship systems, which are organized in complex and varying manners, seem to exhibit an almost optimal tradeoff between simplicity and informativeness (Kemp & Regier 2012). However, in contrast to the location concepts, kinship terms have multiple vectors that could be sound symbolically encoded. Hence, the main criterion used to select these concepts was to capture as many kinship terms as possible and included all blood relations two steps from the ego, with relative age distinctions when applicable, e. g. younger sister’s son, while excluding more distant relations, non-blood relations and umbrella terms, e. g. grandparent and sibling. For a complete list of the 64 selected kinship concepts see Table 2.

Body part concepts are perhaps some of the most fundamental linguistic concepts, but the linguistic segmentation of the body is highly language-specific. While it is easy to assume that body part nomenclature is primarily determined by visual features, there is evidence that proprioceptive (Enfield et al. 2006), developmental (Andersen 1978), and neurological (Penfield & Boldrey 1937; Penfield & Rasmussen 1950) factors also make important contributions. Furthermore, it has been proposed that most languages adhere to a possibly universal hierarchy of lexicalized body parts (Andersen 1978), for the most part corroborated by the fact that joints act as boundaries between body parts in distance judgements (Enfield et al. 2006; de Vignemont et al. 2009). Thus, body part concepts considered fundamental according to these criteria were included (arm, back, body, breast, chest/trunk, ear, eye, face, finger, foot, hand, head, leg, mouth, neck, nose, toe). However, chest/trunk was replaced by the more distinctive belly, face was excluded in favor of head, and arm and leg were further divided into upper arm, lower arm, thigh and lower leg. In addition, body part concepts with distinctive appearances and/or many nerve endings were included (buttocks, fingernail, hair, navel, tooth and skin, lip, throat, tongue, penis, vulva, nipples, testicles), as well as the most distinctive internal organs: heart, lungs and brain. We also included all salient bodily and mental functions related to eyes (cry), mouth (bite, blow, breathe, cough, drink, eat, laugh, say, snore, spit, suck, vomit, yawn), nose (sneeze), genitals/excrement (defecate, intercourse, semen, urinate), skin (blood, milk, sweat), mind (know, sleep, think), movement (fall, go, lie, run, sit, stand, turn) and living (die, live). Generally, the verb related to the function was chosen, except for blood, milk, semen and sweat. hiccup, burp and menstruate were excluded due their similarities to cough and blood, respectively.

According to several studies (Viberg 1983, Viberg 2001), sensory concepts are hierarchically lexicalized, sight being the most fundamental, possibly because there are more occasions to talk about visual objects than about objects related to other senses. Furthermore, touch, taste and smell words often lexically overlap with other senses (San Roque et al. 2015), and not much work has been done on the sound symbolic aspects of these concepts. Accordingly, all typical sense words were selected (see, hear, taste, touch, smell). Similarly, despite criticism of the four traditional basic taste distinctions (Erickson 2008) and various lexical conflations of taste terms occurring throughout languages, cross-linguistic data does support bitter, salty, sour and sweet as fundamental taste concepts (Majid & Levinson 2008), which were therefore included in the list. Basic emotion concepts were, on the other hand, somewhat generalized following Jack et al. (2014), resulting in happy, sad, afraid and angry.

We also selected a number of natural entities that would be salient features in the surrounding world of pre-agrarian societies (bone, fire and sand), general plant and animal concepts (dog), as well as concepts relating to weather, heaven (e. g. sky and sun), (bodies of) water, day and night, but not ice and snow, as they are unknown in many parts of the world. In addition, in order to make the sample of concepts comparable to many other studies which incorporate basic vocabulary and to estimate base frequencies of each sound, it is crucial to include a substantial number of concepts which are likely not affected by sound symbolism. We therefore also added the remaining concepts present in the Swadesh-100 and Swadesh-207 lists (Swadesh 1971), the Leipzig-Jakarta list (Haspelmath & Tadmor 2009) and Goddard and Wierzbicka’s (2002) semantic primes. Altogether, this included air, animal, ant, ashes, bark, because, bird, blunt, bone, burn, carry, clean, cloud, come, correct, crooked, crush, dirty, do, dog, dust, earth, egg, fart, feather, fire, fish, flesh, flower, fly (n), fly (v), give, grass, grease, grow, half, hide, hit, horn, house, if, kill, knee, leaf, liver, loud, louse, maybe, moon, mountain, name, not, part, path, person, pointy, quiet, rain, raw, ripe, river, root, rope, rotten, sand, sea, seed, shadow, sharp, sky, smoke, star, stone, straight, strong, sun, swim, tail, take, thunder, tie, tree, want, water, weak, wind, wing, word, wrong, year and yesterday.

3.2 Capturing linguistic diversity without genetic bias

Undoubtedly, simulating the diversity of human language as a whole by selecting a number of spoken languages is a complicated matter. Not only are languages incredibly diverse in terms of phonology, morphology, lexicon, semantics and syntax, but they also differ widely in the number of speakers, geographical spread and in the number of genetic relatives. However, the main point of selecting a sample of languages is to represent diversity and, for this particular research question, lexical diversity. As relations between languages are largely determined based on lexical differences, a more cautious approach for grouping languages into families is preferable. Therefore, Glottolog’s (Hammarström et al. 2017) approach was adopted in favor of the other currently largest language database, Ethnologue (Simons & Fennig 2017), except in the few cases when Ethnologue’s language division was more conservative. Even if complete datasets for all the world’s documented languages were available, we would still not get the complete picture of what human language is capable of, as most languages are already dead.

The aim of sampling is therefore to include one representative from all the world’s living and extinct documented language families (and isolates) with sufficient and reliable data for at least one member, spread geographically as widely as data availability allowed. In addition, this also compensates for the concepts that lacked data for some languages, since the language sample remains genetically balanced regardless of the number of included languages. Thus, after excluding artificial, sign, unattested and unclassifiable languages, as well as creoles, mixed languages, pidgins and speech registers, because they are mostly based on already existing languages, 245 languages and language families were selected (58.5% of the 419 featured on Glottolog), of which 68 were isolates (Figure 1 and Online Appendix 2). This sample of languages yielded almost 70,000 lexemes.

Figure 1:

Featured languages divided into the same geographical macro-areas as used by Blasi et al. (2016) and shown by color (olive green: North America, blue: South America, forest green: Eurasia, purple: Africa, orange: Papunesia, pink: Australia).

3.3 Data collection

One of the challenges of compiling cross-linguistic data is data collection. For languages with many speakers or long histories as literary languages, comprehensive and reliable sources such as databases or comprehensive dictionaries make the collection of data straightforward. For many poorly documented languages, on the other hand, only a handful of sources have ever been produced, and usually only one or two of those are available. Data availability was thus an important consideration guiding language sampling. Furthermore, due to the varying quality of data, some concepts were not retrieved from all languages, but since only one language per language family was included, the sample remained unbiased.

Moreover, even when obstacles related to data availability have been overcome, differences between languages, such as grammatical marking, still pose problems. For example, when concepts were found to have multiple forms (e. g. gender inflections), only the unmarked form was selected to ensure comparability across languages, as long as relevant information about the meaning was provided through the lexical entries or grammatical descriptions, i. e. in the singular nominative for accusative systems, in the singular absolutive for ergative systems, and so forth. In many languages, the same concept can have a number of different roots or versions, e. g. in classificatory verbs in native North American languages (Kibrik 2012), which makes it difficult to know which form of a group of words is the unmarked one. Likewise, throughout languages, most concepts also have several synonyms. Therefore, all phonemes from all forms in these cases were combined into a single string rather than selecting only one of the forms to represent the concept in question. For example, the three English forms of the third person singular personal pronoun (he, she and it) were analyzed as a single word with six phonemes [hi:ʃi:ɪt]. Conversely, when the same term is used for more than one concept, both slots were filled with the same form. For example, in Pirahã [myp] ‘I’ and ‘we’ are both referred to as ti³ . In addition, large bodies of water are of great import for all speech communities, and thus the concept sea naturally belongs in the list of featured concepts. However, since many cultures lack contact with oceans and thus have no specific word referring to sea, in these cases lake was added instead. This was the only replacement of this kind.

Although including borrowed linguistic forms in cross-linguistic comparisons might seem counterintuitive, it does not by default result in an areal bias. A description of a language is only a snapshot of an ever-changing dynamic system, which means that if a word is borrowed and used, it is also part of the language. And in time, the borrowed words usually adapt to the semantic and phonological framework of the language. However, detected late loans from languages with a strong influence on other cultures, namely Arabic [ara], English, French [fra], Malay [msa], Mandarin Chinese [cmn], Portuguese [por] and Spanish [spa], were removed since the same loans from these languages often occur in a great number of languages and could therefore be mistaken for overrepresentations of sounds. Among these loans, we find e. g. penis, milk, salt, numerals, several color concepts, animal, body, if and year. All featured languages with large Sino-Xenic vocabularies have different lexical registers, and thus native words for a concept were selected when available. In the cases without native forms, many of the loans were kept as the vast majority were borrowed more than eight centuries ago and have undergone extensive phonological and often semantic change, unless the linguistic form showed considerable similarity across the Sino-Xenic languages. Likewise, loan words from less culturally influential languages that were only found in one target language were kept when no native form was found, especially if the word was borrowed within a language family.

3.4 Data transcription model

The inconsistency, quality and granularity of the sources also cause ripple effects for transcription of the collected data. Furthermore, the poor quality of many orthographies, especially of less studied languages, combined with the fact that different sources describing the same language often use different kinds of orthographies and are frequently based on the mother tongue of the data compiler, adds to the overall disarray. In other words, it is nearly impossible to make a dataset with a larger number of languages completely comparable without employing a unifying transcription system. Therefore, all sounds were transcribed into The International Phonetic Alphabet as accurately as the sources for the featured languages allowed for, albeit with some minor yet crucial differences. After the lexemes had been collected, we obtained phonological and orthographical descriptions from the same sources when available in order to convert the text into IPA. When this information was not available, which was the case for several older sources, we consulted available grammatical and phonological sketches and articles. We also utilized phonological data from databases and compilations of phoneme inventories, such as PHOIBLE (Moran et al. 2014), that described the languages in question. As for the featured extinct languages, most of them went extinct in recent times and are therefore quite well-described. However, the few included ancient languages, such as Sumerian [sux], obviously entail more phonetic uncertainty despite the amount of research that has gone into describing them.

The main aim of the current paper requires a quantification and statistical measuring of sound symbolic associations from a cross-linguistic perspective. This aim demands a model of data transcription that is capable of 1) capturing the diversity of various phonemic systems, 2) quantifying these diverse systems in a manner that is representative, comparable, and relevant to the research theme of the paper, sound symbolism. While IPA provides a detailed description of speech sounds, it can be too fine-grained for comparing such a diverse range of languages with highly dissimilar sound systems. Therefore, some sounds needed to be grouped together or segmented in order to make them statistically analyzable. In addition, these classifications should also correspond to how features of speech sounds are observed to behave with respect to sound symbolic mappings in languages. To begin with, all original IPA oral and nasal vowels were included, as well as all pulmonic, doubly articulated consonants and consonants with secondary articulation and non-pulmonic consonants. Voicing was also distinguished by contrasting complete voicelessness with all degrees of voicing, i. e. also including partial, weak and short voicing (Cho & Ladefoged 1999), in accordance with how voicing is mapped sound symbolically (Lockwood et al. 2016a).

Sounds that incorporate more than one place of articulation were split into two segments in order to quantify them separately. This for several reasons: a labialized velar stop, [kʷ], might be used sound symbolically to indicate abruptness through the stop, or to indicate a round shape through the rounding of the lips. Thus, it cannot be equated only with [k] or [w] since the other feature would have been left unnoticed in the data. This model, which is based on how sound symbolism is observed to be reflected in language, may vary with respect to how precisely the phonemic systems of languages are rendered (some languages may have richer systems). However, the model captures the crucial acoustic features important for sound symbolism and allows these features (which are partly phonemes, partly acoustic representations) to be grouped in a more appropriate way than dedicated labels for combinations of phonemes. Hence, diphthongs and triphthongs were transcribed as sequences of vowels, and affricates as combinations of plosives and fricatives because of the shared closure phase between affricates and plosives, and the shared friction phase of affricates and fricatives (Sidhu & Pexman 2018).

Furthermore, the meanings that are sound symbolically associated with affricates are usually semantically similar to both meanings associated with stops and meanings associated with fricatives (Abelin 1999: 37–41). The same principle was applied to ejective affricates and consonants with double and secondary articulation, such as consonantal release types, as well as consonants with aspiration (including preaspiration), labialization and palatalization. For example, [tsʼ], [k͡p], [pᵐ], [kʰ], [kʷ] and [kʲ] were transcribed as /tʼsʼ/, /kp/, /pm/, /kh/, /kw/ and /kj/, respectively. In contrast, breathy (murmured) vowels and nasalized and creaky voiced sounds were coded as separate phonemes since the involved features are difficult to distinguish. Plain click consonants, such as [ʘ], were considered voiceless and contrasted with voiced variants such as [ᶢʘ]. While aspirated and glottalized click consonants were segmented as described above, nasalized clicks and voiceless nasalized clicks were considered separate phonemes, and clicks with a velar, velar ejective, uvular and uvular ejective fricative release were transcribed as a click followed by /x/, /xʼ/, /q/ and /qʼ/. Stress and tones were not recorded since information about stress patterns was generally lacking or poorly described for most languages, and tones occurred only in a fraction of the language sample, which would lead to very low comparability.

Phonetic length was recorded in the form of a double occurrence of the same phoneme: for example, [a:] resulted in /aa/. While this is a simplification, it does retain the perceptual length, which languages with long vowels in their phonological systems could utilize for either quantitative iconicity or for emphasizing sound symbolic segments grounded in qualitative iconicity. Coding long and short segments (such as [a] and [a:]) as different sounds would, on the other hand, fail to record the qualitative similarity between them (for example [a] and [a:] coded as /a/ and /a:/), and coding them as the same sound (for example [a] and [a:] both coded as /a/) would not record potential qualitative iconicity.

3.5 Phonetic categorization

Sound-meaning associations are seldom restricted to one specific phoneme; sounds with similar phonetic characteristics are often used for the same meaning in different languages depending solely on what sounds are accessible for the languages in question. If the purpose is to find statistical evidence for cross-linguistic sound-meaning associations, it would seem unwise to count a plain bilabial [m], a creaky voice bilabial [m̰] and a plain labiodental [ɱ] as separate phonemes due to their phonetic similarity. In addition, sound symbolic associations may not necessarily be grounded in phonemes as such, but rather in the acoustic and/or motor features that define them (Sidhu & Pexman 2018). For example, an association between [m] and mother might only be based on its nasal, and not labial, quality. Thus, as this element has not been incorporated in previous studies, it is crucial to systematically group phonetic parameters to pinpoint the features responsible for each sound symbolic association. There are various ways of analyzing and grouping the features of human speech sounds, but cross-linguistic frequencies of sounds as well as phonetic and phonological similarity are generally the most informative parameters that can be used for this purpose (Mielke 2012). Similarly, sounds can be reduced to a set of distinctive features which can be used to describe most sound classes (Mielke 2008). However, while most of these distinctions are appropriate for describing languages phonetically and phonologically, several distinctions are not relevant for studying cross-linguistic sound symbolism and in some cases can even muddy statistical analyses. For example, typologically uncommon distinctions are by definition difficult to compare across languages, but more importantly, several distinctive features sometimes have to be grouped in order to expose sound symbolic relationships caused by a more general feature. Therefore, all human speech sounds were grouped according to salient articulatory parameters in conjunction with distinctive acoustic features which have been shown to evoke sound symbolic associations in experimental and cross-linguistic studies.

3.6 Data analysis

The goal of data analysis was to identify words with over-represented sound groups – for example, words that contain an unexpectedly high proportion of high vowels across the sampled languages. We started with the assumption that each language has a typical distribution of vowels by height (and other features of interest listed in Table 3) and estimated this distribution by looking at all 344 sampled words from that language. If, in many of the sampled languages, a particular word contained a markedly higher proportion of rounded vowels than the average for each language, we interpreted this as evidence that some force, such as sound symbolism, was driving this non-arbitrary word form.

Calculating the absolute number of phonemes occurring within a word could skew the results through, for example, reduplication and effects of word length. Previous comparable studies did not include concepts which often involve reduplication (kinship concepts, numerals, etc.); hence, reduplication and similar phenomena were not controlled for, even though these phenomena also affect a range of other basic vocabulary items. Furthermore, the aim of this study was to investigate the occurrence of phonemes across languages, not their occurrence within specific linguistic forms. To avoid this problem, we chose to analyze proportions rather than absolute counts of sound groups. These proportions were calculated separately for vowels and consonants. So, for example, the word /mantu/ (“belly” in the Ngarinyin language [ung]) contains 66% of voiced and 33% of unvoiced consonants; 50% of high, 0% of mid, and 50% of low vowels; and so on. A hypothetical complete reduplication into /mantu-mantu/ would have no effect on these proportions, and it would therefore remain “invisible” to the model. In contrast, a partial reduplication of one syllable (e. g. /mantu-tu/) would affect the proportions of sound groups. Likewise, because long vowels and diphthongs were coded as two separate phonemes (e. g. /ma:ntu/ would be coded as /maantu/), sound symbolic prolongation of vowels was captured by the models we used. As a “bonus”, this approach also solves the problem of some concepts being represented by more than one linguistic form, e. g. the English he, she, it.

A transformed dataset of proportions was prepared and modeled separately for each of 10 evaluated sound groups: backness, height, roundedness, extreme and extreme-roundedness for vowels; manner, manner-voicing, place, place-voicing and voicing for consonants (Table 3). One row in the dataset corresponded to one word in one language, and the response variable was a vector of proportions that summed to one – in mathematical terms, a simplex. We modeled these simplex responses with the Dirichlet distribution in the framework of Bayesian generalized linear models (GLM) as implemented in the R package brms version 2.9.0 (Buerkner 2017), with default conservative priors.

Using vowel height as an example, the model included a population-level intercept corresponding to the overall distribution of vowels by height across all words and languages, a group-level (random) intercept per language corresponding to the typical distribution of high, low and mid vowels in each particular language, and a group-level (random) intercept per word. This random intercept per word was the measure of interest, since it showed deviations from the typical distribution of vowels by height in particular words. As usual with multilevel models, representing proportions for each word and language as drawn from a single distribution imposed shrinkage – that is, drew the estimates closer to the group mean. The amount of shrinkage was controlled adaptively by the data itself, which is a great advantage of multilevel models and the reason why the effect of word was modeled as a random rather than fixed effect. Shrinkage was stronger when the outcome variable had many levels and more moderate for outcomes with two levels, such as voicing and roundedness; it was also stronger for rare sound groups with relatively few observations (e. g. voiced glottals), where the apparent outliers were driven by only a few languages (Online Appendix 3).

The output of interest from these Dirichlet models was a list of fitted proportions of sound groups (e. g. of high, low and mid vowels) in each of 344 words. To identify cases of over- or underrepresentation, we also extracted fitted average proportions of each sound group (e. g. high vowels) across all words and then compared per-word estimates to these average values. To propagate uncertainty of model estimates, this comparison was performed for each step in the Markov chain Monte Carlo, resulting in a posterior distribution of how much each word deviated from the typical distribution of sound groups.

One way to compare distributions would be to look at simple differences of proportions of each class. For example, if the typical proportion of high vowels is 50% and a word contains (on average across all languages) 55% of high vowels, this constitutes a 5% overrepresentation. The problem with this approach is that it does not scale very well for proportions that are close to 0% or 100%. For example, if the frequency of a rare sound group jumps from a base rate of 5% to 10% in a particular word, this is substantively a greater change than from 50% to 55%. To account for this, we compared odds ratios (OR): an increase from 5% to 10% corresponds to OR = 1:9/1:19 = 2.1, while an increase from 50% to 55% gives OR = 11:9/1:1 = 1.2.

Since we employed a Bayesian analysis, we did not test the statistical significance of any effects. Instead, we defined a region of practical equivalence (ROPE), symmetric on a logarithmic scale, around the null effect of no overrepresentation (log-odds ratio = 0 or, equivalently, OR = 1). The width of the ROPE corresponded to a change of OR by a factor 1.25 1 * 1.25 = 1.25, or +25%; 1/1.25 = 0.8, or −20%). This ROPE was set to represent the smallest substantively interesting effect size: a 25% increase of OR corresponds to an increase in the proportion of a sound in a word from 10% to 12%, 50% to 55.5%, 90% to 92%, etc. Following the guidelines for decision making in this analytical framework (Kruschke & Liddell 2018), we distinguished between three types of outcome:

1. “Strong association”: if the 95% credible interval (CI) for the OR fell completely outside the ROPE, we concluded that the distribution of sound group in this word substantively deviated from the distribution expected by chance.

2. “No association”: if the 95% CI was completely contained inside the ROPE, we concluded that there was no over- or underrepresentation.

3. If the 95% CI partly overlapped with the ROPE, the result was treated as ambiguous. Because there was a substantial number (~9%) of such cases, we further distinguished between two subtypes. If the 95% CI excluded zero and the median of posterior distribution (our “best guess”) was outside the ROPE, the association was treated as “weak” but potentially interesting; otherwise it was treated as too uncertain for being considered further.

It is worth emphasizing that the ROPEs refer to fitted rather than observed values. In most models and categories, shrinkage of regression coefficients to zero was very pronounced (see Online Appendix 3), thus producing very conservative estimates of the degree of under- or overrepresentation. As a result, the number of associations reported below (225, or ~1.3%) is vastly lower than the number of cases for which the observed OR lies outside the same ROPE (6708, or ~36% of all possible associations).

4.1 General results

The total number of potential associations was very large, varying from 344 in models with two sound groups (e. g. voiced or unvoiced consonants, rounded or unrounded vowels) to 3096 in the models with ten sound groups (Place-voicing and Manner-voicing), for a total of 17,888 possible associations across ten models. However, an overwhelming majority of associations was classified as absent (90.8%) or doubtful (7.9%), leaving only 176 (1.0%) weak and 49 (0.3%) strong associations (Figures 2 and 3). These numbers exclude cases of underrepresentation of sound groups with two levels (vowel roundedness and consonant voicing) since these were redundant mirror images of overrepresentations. For example, if rounded vowels are overrepresented in a particular word, unrounded vowels register as equally underrepresented. Cases of underrepresentation could be of some interest for oppositional concepts of binary or continuous domains. For example, sounds overrepresented in big might be underrepresented in its opposite small to emphasize the contrast. However, the results yielded few clear sound symbolic antonyms, making the negative associations difficult to interpret: there could be many reasons why some sounds seldom occur in a specific word. In some cases, underrepresented sounds could be a consequence of other classes being strongly overrepresented, particularly in short words. However, we did not find any correlation between word length and the probability of a word being sound symbolically affected. Thus, we only focus on the overrepresented associations in the following discussion.

Figure 2:

Over- and underrepresented sound groups in 344 concepts: strong (black) and weak (gray) associations. Each point shows the median of posterior distribution of the ratio of observed to expected odds, with 95% CI. Text labels show the concept, associated sound group, and the most strongly over- or underrepresented cardinal sound in parentheses. The marked region of practical equivalence (ROPE) of [0.8, 1.25] was used to select substantively relevant findings.

Figure 3:

Over- and underrepresented sound groups in 344 concepts: strong (black) and weak (gray) associations. Each point shows the median of posterior distribution of the ratio of observed to expected odds, with 95% CI. Text labels show the concept, associated sound group, and the most strongly over- or underrepresented cardinal sound in parentheses. The marked region of practical equivalence (ROPE) of [0.8, 1.25] was used to select substantively relevant findings.

If we compare the found associations with previous similar studies, it becomes evident that the investigated concepts have varied considerably, see Table 4. For the present study, the associated sound group with the highest specificity is listed, e. g. if a concept was associated with [high], [high-back] and [high-back, +r], only [high-back, +r] was listed. 19 concepts and 20 associations (ashes-[back], bone-[–voice], breast-[nas, +v], F_fs/F_ms-[lab]/[low-front, –r], 1sg-[nas, +v]/[–round], knee-[+round], M_fs/M_ms-[nas, +v], nose-[nas, +v], skin-[–voice], tongue-[alv, +v], 1pli/1ple-[nas, +v], 2sg-[nas, +v]/[–round]) also clearly correlate with those reported by Johansson (2017), Wichmann et al. (2010) and Blasi et al. (2016) and therefore ought to be considered very robust. In addition, several of the associations were also found to be similar to previous findings. For example, hard was found to be associated with voiceless alveolars by Johansson (2017) and with (mostly voiceless) stops in the present paper. Likewise, while short and small were found to be associated with voiceless alveolars, /i/ and /C/ by Johansson (2017) and Blasi et al. (2016) but with [stop, –v] (as well as [rounded]) in the present results, all sound groups involve high frequency sounds. Furthermore, the present study found another 39 concepts and 105 associations which are described in Section 4.2. There were, however, also several discrepancies between the present and previous studies. Johansson (2017) found several associations to sound groups that generally contain few sounds, i. e. vibrants, laterals and voiced palatals in deep, flat, hard and this. This is likely a result of the considerably smaller and less balanced sample of languages and the less robust statistical analysis. It is possible that the overrepresentation of voiceless labials in flat found by the same study is a similar case. Both Blasi et al. (2016) and Johansson (2017) also found associations between round and vibrants, while the present study found associations to [back] (as well as [+round]), mainly represented by /u/. The association between round and rounded sounds is further discussed in Section 4.2.2 but is rather straightforward to understand. However, the “lack” of overrepresentation of vibrants could be attributed to the strict modeling used in the present study which might have created a higher confirmation threshold for the investigated sound-meaning associations compared to previous studies. Our cautious approach could also therefore have resulted in the loss of several potential associations. What is more, the semantically similar concept turn was found to be associated with [alv, +v] (mainly represented by /r/), which also suggests a connection between circular shapes and vibrants. A full list of all associations and concepts is found in Online Appendix 1 along with cardinal sounds, overall occurrence, as well as the type of sound symbolic mapping and associated macro-concept, as explained below.

4.2 Macro-concepts based on semantic and phonetic common denominators

Overall, all of the discovered sound-meaning associations belonged to bodily functions, body parts, deixis, descriptors, kinship terms, logical concepts, or natural entities. More interestingly, however, the concepts with noteworthy overrepresentations could in turn be grouped into semantically and sometimes phonetically superordinate concepts, here referred to as macro-concepts. Arranging the discovered associations in this manner has several benefits: a) it provides an overview of the rather long list of confirmed sound-meaning associations in an exploratory study such as the present one, and b) it makes it possible to use observable semantic and phonetic regularities to further understand how sound symbolism could be used to define fundamental lexical fields in human language. The macro-concepts should therefore be regarded as preliminary classifications, but could still act as a stepping stone for future studies. This grouping required that the confirmed sound-meaning associations shared both semantic and phonetic features and was defined as follows.

Strong macro-concepts had to include at least one of the strong sound-meaning associations or at least two weak sound-meaning associations. For strong macro-concepts consisting of more than one sound-meaning association, the included associations also had to share one or more concepts that share at least one semantic feature and one or more concepts that share at least one associated sound.

Weak macro-concepts had to include one of the weak sound-meaning associations which could be corroborated by a qualitatively parallel association or macro-concept (e. g. associations between large and low-frequency sounds, and small and high-frequency sounds) or by a plausible sound symbolic explanation in line with known associations reported in other studies on sound symbolism and iconicity. When evaluating shared sound-meaning associations, the most commonly occurring cardinal sounds were taken into account as these are informative in regard to the driving factors behind associations. For example, the effect of an association between a concept and [stop, –v] and [lab, –v] could be driven by an intersecting /p/ in both cases.

This further means that a concept, particularly concepts associated with more than one sound group, can belong to several macro-concepts, and macro-concepts can include various sound groups as long as those sound groups share relevant phonetic features. In addition, the interaction between semantic and phonetic features, as well as cardinal sounds, also makes it possible to trace which type of sound symbolic mapping grounded each sound-meaning association. As the study was designed to be explorative, all possible types were of interest. However, basing the calculated results on relative frequencies of sounds washed away internal word structure patterns, making it impossible to analyze gestalt iconicity, i. e. cross-modal, iconic or indexical mappings of word-internal structural emergence. For example, reduplication occurs frequently in some languages but is almost absent in others. To complicate things even more, words can be reduplicated either completely (e. g. Basque [eus] zapla-zapla ‘slap’) or partially (e. g. Pangasinan [pag] toó ‘man’ and totóo ‘people’). Phenomena such as phonesthemes, i. e. associative, cross-modal, indexical mappings of language-internal analogical emergence, also had to be excluded since they are not detectable due to their language-specific character.

A complete list of all macro-concepts, their contained concepts and associated sound groups, as well as the most frequently occurring cardinal sounds in each sound group association and sound symbolic mapping types, are provided in Table 5.

In total, the results revealed 134 sound-meaning associations. We did not find any plausible explanation for the associations between back and [+round], empty and [+round], think and [nas, +v] and tie and [–voice], while those between blow and [central] and suck and [central] were only found in 11 and 14 languages, respectively. Therefore, these associations were judged as doubtful. Furthermore, short was unexpectedly associated with [+round] which would be the reverse of the expected pattern of ‘small’-high frequency and ‘large’-low frequency (see Section 4.2.3). However, as this association does not correlate with any previous findings, it is quite possible that it is a result of noise in the source materials. This association was thus also judged as doubtful. These were therefore excluded from further analysis, resulting in a grand total of 125 relevant associations involving 59 concepts. In addition, since an association can be grounded in more than one way simultaneously, e. g. both through visual and acoustic motivations, there were in total 140 sound symbolic motivations. In turn, these motivations were found across four types of mappings, of which two, vocal gestures and circumstantial mappings, have not previously been explicitly described in the sound symbolic literature (summarized in Figure 4).

Figure 4:

Illustrated simplifications of types of sound symbolism described by Dingemanse (2011), Carling and Johansson (2014) and the present paper. a) Onomatopoeia (imitative): acoustic approximations using the human vocal apparatus. b) Vocal gestures (imitative): cross-modal imitations in which the acoustic signals are only accompanying the gesture. c) Gestalt (diagrammatic): mappings between event structures and word structures, e. g. Swahili piga ‘to strike’ and its reduplicated form piga-piga ‘to strike repeatedly’. d) Relative (diagrammatic): relational mappings between semantic and phonetic scales or poles. e) Complex (associative): language-internal mappings which emerge through analogy. f) Circumstantial (associative): mappings based on circumstantial associations between referents which are part of an event and sounds which are frequently expressed during the same event.

Summarized, of the 140 motivations, 37 (26.4%) were defined as onomatopoeia, 31 (22.1%) as vocal gestures, 16 (11.4%) as relative, 57 (40.7%) as circumstantial and 7 (5%) remain doubtful. Furthermore, macro-concepts consisting of a single concept could in fact be members of yet undefined larger macro-concepts that remain opaque since they include concepts not featured in the present sample.