When shown the shapes in Fig. 1, roughly 90% of individuals worldwide (Styles & Gawne, 2017; see this paper also for exceptions) associate the round one with nonword labels like maluma and the spiky one with nonword labels like takete. This maluma/takete effect (Köhler, 1929) is one example of sound symbolism, associations between formal language sounds and perceptual and/or semantic properties.Footnote 1 These associations may derive from perceptuomotor analogies between language sounds and properties in other modalities (e.g., between the abrupt sounds in takete and the abrupt changes in outline of the spiky shape; see Sidhu & Pexman, 2018). The maluma/takete effect has been observed in speakers of different languages (Styles & Gawne, 2017) and across ages (for a review see Fort et al., 2018). It has been demonstrated using both behavioural tasks (e.g., Nielsen & Rendall, 2011) and neuroimaging (e.g., Asano et al., 2015).

Fig. 1
figure 1

Most people pair maluma with the left shape and takete with the right shape

Full size image

The maluma/takete effect generalizes to several categories of phonemes. The most robust associations are between sonorants (e.g., /l/, /m/, /n/) and round shapes, and voiceless stops (e.g., /p/, /t/, /k/) and spiky shapes (e.g., McCormick, Kim, List, & Nygaard, 2015). Some have also found associations between voiced stops (e.g., /b/, /d/, /g/; most commonly /b/) and round shapes (e.g., McCormick et al., 2015). In addition, studies have shown associations of rounded back vowels (e.g., /u/ as in boot) with round shapes, and unrounded front vowels (e.g., /i/ as in beet) with spiky shapes (D’Onofrio, 2013; McCormick et al., 2015). In the largest-scale study to date, Westbury, Hollis, Sidhu, and Pexman (2018) examined the fit between 8,000 nonwords and roundness or spikiness. They found that the phonemes /oʊ/ (as in boat), /u/, /b/, /m/, and /ɑ/ (as in bought) were associated with roundness, and that the phonemes /t/, /k/, /z/, /i/, and /ɪ/ (as in bit) were associated with spikiness.Footnote 2

Interest in sound symbolism is motivated in part by the long-debated relationship between linguistic form and meaning. The dominant view has been that this relationship is arbitrary (e.g., Hockett, 1963). One possible opposition to this is iconicity. In spoken language iconicity here refers to instances in which aspects of a word’s form somehow resemble aspects of its meaning (Dingemanse, Blasi, Lupyan, Christiansen, & Monaghan, 2015).Footnote 3 The most obvious examples are when the sound of a word imitates a sound-based meaning (i.e., onomatopoeia; e.g., the sound of quack resembles duck sounds). However, as discussed, language sounds can also have associations with nonauditory properties (e.g., shape). This allows for the possibility of cross-modal iconicity. For instance, the phonemes in balloon are associated with roundness, allowing balloon to resemble its meaning via the associations of its component phonemes, making it iconic (see Table 1 for definitions).

Table 1 The terms sound symbolism and iconicity, as they will be used in this paper
Full size table

While sound symbolic associations have been demonstrated using laboratory tasks with nonword stimuli, the presence of these associations in real language has not been established. That is, it is unknown whether there is a tendency for languages to contain cross-modally iconic words like balloon. Most investigations of cross-modal iconicity in real language have focused on a small number of words.Footnote 4 Many have investigated size iconicity. That is, whether the sound symbolic association between high-front vowels (e.g., /i/ as in beet) and small shapes, and low-back vowels (e.g., /ɑ/ as in bought) and large shapes (Newman, 1933; Sapir, 1929), is present in real langauge. One approach has been to analyze a sample of words in a language and see whether words for small or large things contain small-associated or large-associated phonemes. The results of this approach have been equivocal. Some studies find evidence of size iconicity in English (Thorndike, 1945). Some do not (Katz, 1986; Newman, 1933). Taking another approach, Blasi, Wichmann, Hammarström, Stadler, and Christiansen (2016) investigated the forms of 100 basic words across two-thirds of the world’s languages. They found a tendency for the word meaning small to contain the phoneme /i/.

There has been less work exploring shape iconicity (i.e., the maluma/takete effect in language). Blasi et al. (2016) found that words meaning round tended to contain the phoneme /r/. They (and others; Johansson et al., 2020; Joo, 2020) have also found that across languages, words for some round body parts tend to contain round-associated phonemes (e.g., words for breast containing /m/, /u/). Katz (1986) explored the vowels contained in 325 English concrete words that had been rated on their shape (from round to angular). There was a trend in which words containing /u/ were the most round, though the overall effect of vowel type was not significant. More recently, Monaghan, Mattock, and Walker (2012) examined English words relating to roundness and angularity (311 and 198 words, respectively) Words for angularity were more likely to contain velars (e.g., /k/, /g/) and unvoiced consonants (e.g., /p/, /f/), but not after correcting for multiple comparisons.

Thus, whether shape iconicity is present in large samples of a lexicon is still unknown. This is an important question. Research on the maluma/takete effect in nonwords has proliferated in recent years, yet the relevance of this work to real language has remained an intriguing but unanswered question. In the present study we conducted the first large-scale investigation of shape iconicity in existing language. We first conducted an exploratory analysis to examine whether there are differences in the phonemes that tend to appear in English words for round versus spiky objects. Then, we conducted a confirmatory analysis to directly test whether round-associated phonemes (e.g., /oʊ/, /u/, /b/, /m/, and /ɑ/) are more common in English words referring to round objects, and spiky-associated phonemes (e.g., /t/, /k/, /z/, /i/, and /ɪ/; Westbury et al., 2018) are more common in words referring to spiky objects.

Method

Participants

A total of 171 participants (43 males; Mage = 20.3 years, SD = 3.9; 84 at the University of Alberta and 87 at the University of Calgary) participated in exchange for partial course credit. All participants reported English fluency and normal or corrected-to-normal vision. We did not require that participants were native English speakers.

Materials and procedure

We chose to examine words referring to objects in order to have a large set of items (e.g., larger than if we had examined shape adjectives). We began with a list of more than 8,000 nouns in the CELEX database that had standardized concreteness ratings > 1.5 in Hollis, Westbury, and Lefsrud (2017). Hollis et al. derived these concreteness ratings by statistically extrapolating from 37,058 human-rated words (human ratings from Brysbaert, Warriner, & Kuperman, 2014, using a 5-point scale ranging from abstract to concrete) to a dictionary of 78,286 words, from a model that used Word2Vec vector values as predictors. We then conducted a pilot study in which a group of 67 participants (who did not participate in any other studies reported here) categorized subsets of these words as an “object” (e.g., blanket, flag, mailbox), “not an object” (e.g., barber, herd, paisley) or as an unknown word. We retained words that a majority classified as an object. We removed plurals and words referring to an object whose shape would be difficult to rate (e.g., mass nouns and objects without a defined shape, such as veggie or gizmo), to end up with 1,757 singular object nouns.

To obtain shape ratings of these objects we made use of best–worst ratings, which are more reliable and efficient than rating scales for collecting semantic norms (see Hollis, 2017; Hollis & Westbury, 2018; Kiritchenko & Mohammad, 2017). Participants were tested individually. Each rated 100 sets of six words. Participants saw six words at a time in the middle of a computer screen, presented using custom-written software. The order of the words, and the order of the 100 word sets, were random. On each trial, participants used the mouse to choose the words referring to the most round and the most spiky objects (see Fig. 2). Using software released by Hollis et al. (2017) for this purpose (available from https://sites.ualberta.ca/~hollis/), we ensured that the trial-to-word ratio was close to 8:1. Trials were constructed to contain minimal informational redundancy. Since every trial involved six words, this means that each word was expected to be judged 8 × 6 = 48 times. Previous work has found that a trial-to-word ratio of 8:1 is sufficient for reliable best–worst judgments (see Hollis, 2017; Hollis, 2019; Hollis & Westbury, 2018). Because we ended up with slightly fewer participants (171) than our goal (174), each word was judged on average 7.9 × 6 = 47.4 times. We used value scoring (explained in Hollis, 2017) to calculate spiky-round scores for each word based on these data (i.e., shape rating). The value-scoring algorithm predicts where an item should be rated relative to other items, based on its history of being chosen as rounder or spikier than other items, adjusting for the competitiveness of the items it appeared with (i.e., how often they were chosen as roundest or spikiest). Table 2 presents the 10 word referents judged as roundest and spikiest (all the ratings can be found at https://osf.io/nfkd2).

Fig. 2
figure 2

An example trial. Participants were presented with six object words at a time and had to choose those that referred to the most round and most spiky objects

Full size image
Table 2 The words judged to have the 10 roundest-shaped and spikiest-shaped referents based on best–worst ratings
Full size table

Results

Our initial exploratory analysis examined which phonemes were predictive of a word referent’s shape. Our second analysis was confirmatory, using existing models of phonemes’ associations with roundness and spikiness (Westbury et al., 2018) to quantify the sound symbolic associations of each word’s sound, and to determine whether this predicted each word referent’s shape.

Exploratory Analysis

The data were analyzed using least absolute shrinkage selection operator (LASSO) regression (see Friedman, Hastie, & Tibshirani, 2010). This analysis finds regression coefficients that minimize the sum of squared residuals and the sum of the absolute values of all coefficients, by multiplying the summed absolute coefficient values by a value lambda, which is determined via cross-validation and included in the model as error. We used an Adaptive LASSO (using the “glmnet” package in R, Simon, Friedman, Hastie, & Tibshirani, 2011) which tunes the value of lambda for each of the predictors separately (see Zou, 2006). We used a common definition of lambda as the largest value of lambda that results in a model with mean cross-validated error within one standard error of the minimum (Friedman et al., 2010). LASSO models are conservative with regards to large coefficient values and guard against overfitting, with many predictor coefficients being shrunk to zero. Annotated R code can be found at: https://osf.io/nfkd2.

The independent variables of interest were 39 dichotomous predictors that coded for the presence of every phoneme. We coded for the presence of each phoneme (rather than the number of each phoneme), because it was rare for any phoneme to appear more than once in a word. On average, each phoneme appeared more than once in only 0.75% of itemsFootnote 5 (see Fig. 3 for the frequency of each phoneme). We also included log-transformed word frequency +1 (Shaoul & Westbury, 2006), and number of phonemes (M = 5.07, SD = 1.53) as control variables. These variables were standardized and “forced” into the model by setting their lambda values to zero (i.e., they could not be removed by the LASSO process). The dependent variable was each word referent’s shape rating. Thirty-nine words were excluded for not having available frequency values. One word was excluded because it was misspelled which left a total of 1,717 items.

Fig. 3
figure 3

The proportion of words used in the analysis that contained each of the phoneme predictors

Full size image

The resulting model is summarized in Table 3. Nearly all of the phonemes that were more common in words referring to a round object in the present study have been shown in previous studies to have a sound symbolic association with roundness, except for /i/. Three of the phonemes that were more common in words referring to a spiky object (/k/, /t/, /ɪ/) have been shown in previous studies to have a sound symbolic association with spikiness. Additionally, the vowels /aɪ/ (as in bite) and /ɝ/ (as in bird) are consistent with the vowels typically associated with spiky shapes (i.e., front unrounded vowels). We also ran versions of this model only including noncompound nouns (n = 1,296), or monomorphemic words (n = 1,013; Balota et al., 2007). Three predictors (/i/Footnote 6, /ɪ/, /s/) no longer enter these models and thus should be interpreted with caution.

Table 3 Predictors of word referent shape, separated into those predictive of object roundness (positive) and spikiness (negative)
Full size table

We ran analogous analyses predicting valence and arousal (n = 1,318; Warriner, Kuperman, & Brysbaert, 2013), concreteness (n = 1,620; Brysbaert et al., 2014; n = 1,717; Hollis et al., 2017), and size (n = 613; Scott, Keitel, Becirspahic, Yao, & Sereno, 2019) of our items. This was to determine whether phoneme presence would be predictive of any dimension, given a large set of items. No phoneme predictors entered any of these models.

Because results can differ by analysis method, we have included in Fig. 4 results from 12 other approaches to the main analysis (details on these analyses can be found in Electronic Supplementary Material here: https://osf.io/nfkd2). The general pattern is that most predictors of spikiness were robust to different analyses, while predictors of roundness were more variable. Results also differed depending on whether controls (in particular number of phonemes) were included.

Fig. 4
figure 4

Results of 13 different approaches to the main exploratory analysis. Each analysis examined whether the presence of a given phoneme predicted a word’s shape rating. Cells in purple convey that a given phoneme was found to be predictive of a word referring to a round object; those in yellow were predictive of a word referring to a spiky object. These different methods included Adaptive LASSO (as reported in the main analysis), Bonferroni-corrected t tests, multiple regression, stepwise regression including forward and backward changes using Bayesian Information Criterion (BIC), best subsets regression which selected the model of all possible models resulting in the lowest BIC, and two stages of a random forest analysis which calculated the importance of variables across many randomly constructed models (note that the output of these models only indicate whether predictors warrant inclusion in the model, not their coefficients). Full details can be found in Electronic Supplementary Material here: https://osf.io/nfkd2

Full size image

Confirmatory analysis

We began by quantifying the sound symbolic association of each word’s sound based on its component phonemes. This was based on models in Westbury et al. (2018), who asked participants to decide whether 8,000 nonwords were a good label for either an unspecified “round thing” or a “sharp thing” (i.e., on separate trials), and then generated two coefficients for each phoneme reflecting its association with roundness and spikiness. Using these coefficients, we computed the sound symbolic association between each of our words and both roundness and spikiness. This was done by summing the roundness coefficients for each phoneme in a word, and then doing the same for spikiness coefficients (see Fig. S1 for their distribution here: https://osf.io/nfkd2/). Sound symbolic roundness and spikiness were negatively correlated with one another (r = −.52, p < .001).Footnote 7 We summed these two values for each word (after reverse scoring spikiness scores) to create a single sound symbolism score. This quantifies the extent to which the sound of each word is associated with roundness or spikiness (see Table 4).

Table 4 The 10 words from our stimuli whose sounds are most sound symbolically associated with roundness and spikiness, based on the phoneme-only model in Westbury et al. (2018)
Full size table

We then ran a linear regression which included each word’s sound symbolism score, number of phonemes, and logged frequency (Shaoul & Westbury, 2006) as predictors. All predictors were standardized. The dependent variable was each word referent’s shape rating. This model revealed a significant effect of sound symbolism score (b = 0.03, p < .001; see Table 5). Words with round-associated sounds were more likely to refer to round objects, and words with spiky-associated sounds were more likely to refer to spiky objects. The patterns and significance were virtually unchanged if compound nouns or if multimorphemic words were excluded (in both cases bSound = 0.03, p < .001).

Table 5 Linear model predicting each word’s shape
Full size table

Finally, we examined the extent to which an alignment of sound symbolism score and shape rating (i.e., iconicity) agreed with existing subjective ratings of iconicity. We derived a measure of iconicity by multiplying standardized sound symbolism scores and shape ratings. Thus, items with both a round sound and shape (or spiky sound and shape) received positive values, while those showing a mismatch received negative values. There was a significant positive correlation (r = .15, p < .001) between our derived iconicity measure and existing subjective ratings of iconicity (n = 577; Perry, Perlman, & Lupyan, 2015; Winter, Perlman, Perry, & Lupyan, 2017).

Discussion

Associations between phonemes and shapes (i.e., the maluma/takete effect) have been well demonstrated with nonword stimuli in laboratory tasks. Here, we investigated whether these associations are attested in real language (i.e., whether round-associated or spiky-associated phonemes are more common in words referring to round or spiky objects, respectively), and thus whether shape iconicity is present in language. The exploratory analysis showed that four phonemes with sound symbolic associations to roundness (/u/, /m/, /oʊ/, /b/) were more common in words referring to round objects, and three phonemes with sound symbolic associations to spikiness (/k/, /t/, /ɪ/) were more common in words referring to spiky objects. The phonemes /aɪ/ and /ɝ/ were also more common in words referring to spiky objects, which is consistent with front unrounded vowels’ sound symbolic associations with spikiness. Further, when we directly quantified the sound symbolic association of each word’s phonology (as round-associated or spiky-associated), this predicted the shape of a word’s referent.

Two of the vowel phonemes more common in words referring to round objects (/u/ and /oʊ/) were back and rounded vowels. Front and unrounded vowels (/aɪ/, /ɝ/, and /ɪ/) were more common in words referring to spiky objects. The reverse of this pattern was observed in the phoneme /i/ being more common in words for round objects. Note that this association was not present when compound nouns or multimorphemic items were removed. The consonants that were more common in round objects (/m/ and /b/) were both voiced bilabials. Contrasting them with the consonants more common in spiky objects (e.g., /t/ and /k/) reveals a smoother sound and articulation for /m/ and /b/, both of which could be associated with roundness (and the converse with spikiness) through perceptuomotor analogy.

Unexpectedly, the affricate /tʃ/, fricatives /ʃ/ and /s/, and the approximant /r/,Footnote 8 were all more common in words for spiky objects. Winter (2016) found that the phoneme /r/ was more common in words denoting rough versus smooth textures. Aryani et al. (2018) found that words containing hissing sibilants (e.g., /s/ and /ʃ/) tended to be higher in affective arousal. Both the rough/smooth and excited/calm dimensions could partially overlap with the spiky/round dimension, and thus help explain these results. Notably, these phonemes, along with the other consonants common in words for spiky objects, all involve the tongue in their articulation, while those common in words for round objects are all articulated with the lips. We note the somewhat contradictory finding by Blasi et al. (2016) that /r/ is common in words meaning “round” across languages.

Words referring to spiky objects tended to contain more phonemes. This persisted even after accounting for age of acquisition (Kuperman, Stadthagen-Gonzalez, & Brysbaert, 2012) in a supplementary analysis. Part of the explanation could be that spikier objects in our dataset also tended to be larger (n = 613; r = −.26, p < .001), since words for larger objects tended to contain more phonemes (r = .13, p < .001). Thus, the relationship between length and spikiness could be indicative of an iconic relationship between word length and referent size. A mediational analysis found evidence that size partially mediated the relationship between length and shape (Average Causal Mediation Effect = −0.006, p < .001; Average Direct Effect = −0.04, p < .001). Previous work has also demonstrated that participants will associate longer nonwords with more visually complex objects (Lewis & Frank, 2016). Thus, the relationship between length and spikiness could reflect a greater visual complexity in spikier objects.

Our main finding was that many of the associations between phonemes and shapes found in laboratory tasks are attested in the pairing between sound and meaning in English. Certainly, this is a modest effect—phoneme predictors and sound rating had small coefficients in the exploratory and confirmatory analyses, respectively. Many other factors play larger roles in the form of language. Nevertheless, the presence of shape iconicity was observable in the present analyses. The process by which this occurs is unknown. From an evolutionary standpoint, aspects of language that convey some benefit to processing and learnability should have an advantage and thus survive (Monaghan, Christiansen, & Fitneva, 2011). Iconic forms can be easier to learn and remember (e.g., Imai, Kita, Nagumo, & Okada, 2008; Lockwood, Dingemanse, & Hagoort, 2016), perhaps affording them a slight lexical advantage. Future historical studies will be necessary to understand this process.

The opposite interpretation is that shape iconicity in a language creates the maluma/takete effect observed in laboratory tasks with nonwords (see Taylor, 1963). While it is not possible to rule this out, we believe it is unlikely that the patterns observed here are entirely responsible for effects with nonwords. This is because the maluma/takete effect has been observed in speakers of different languages (see Styles & Gawne, 2017). If patterns in language created those effects, then shape iconicity would have had to emerge accidentally in each of the languages in which the maluma/takete effect has been observed. However, it is possible that patterns in language may strengthen sound symbolic associations, creating a feedback loop (see language specific iconicity in Imai & Kita, 2014). An interesting topic for future research would be to examine whether the extent to which shape iconicity is observed in a language predicts the strength of the maluma/takete effect in its speakers.

Another possibility is that words’ sounds affected best–worst ratings. However, the predictive effect of sound score (i.e., the confirmatory analysis) was diminished for items in the middle two quantiles of shape ratings (i.e., those with more ambiguous shapes; b = 0.00, p = .048) compared with words at the extremes (b = 0.05, p < .001). This is not consistent with sound symbolism driving shape ratings, as these ambiguous shapes should be more susceptible to the effect of sound.

The present results are also indicative of another type of nonarbitrariness, namely systematicity: large-scale patterns in the forms of words belonging to the same syntactic or semantic category (Dingemanse et al., 2015); in this case objects of a certain shape. While systematic patterns need not be iconic, the two are not mutually exclusive. The present pattern is perhaps best described as systematic iconicity: there are patterns in the forms of words belonging to the categories of round and spiky, and the specific nature of those patterns is iconic. Since systematicity tends to be pervasive in a language (Dingemanse et al., 2015), systematic iconicity could represent a means by which iconicity is broadly relevant to a lexicon, beyond specific classes of words (e.g., onomatopoeia).

The maluma/takete effect has been studied for nearly 100 years. In that time, its connection with real language has been largely unexplored. The present results suggest that the maluma/takete effect is attested in the English lexicon, in the form of shape iconicity.