Articulatory features of phonemes pattern to iconic meanings: evidence from cross-linguistic ideophones

2.1 Phonosemantics: the study of sound-meaning correspondences

The subfield of phonosemantics subscribes to a broad hypothesis that “every phoneme is meaning-bearing,” in a word and that meaning “is rooted in its articulation” (Diffloth 1972, 1979; Hamano 1998; see Dingemanse 2018 for review). Phonemic sound-meaning correspondences, henceforth phonosemantic mappings, have been proposed for a number of languages (Akita et al. 2013; Assaneo et al. 2011; Ayalew 2013; Blasi et al. 2016; Hamano 1998; Kwon and Round 2015; Maduka 1988; Oswalt 1994; Waugh 1994). For example, the appearance of /p, b/ in ideophones to do with explosions, expectoration, or releases of pressure is explained through the articulatory properties of /p, b/ themselves (a blockage of airflow then followed by a release) which gesturally resemble those meanings. In the present study, even though we do not assume that absolutely all phonemes are necessarily meaning-bearing in all contexts, we do subscribe to the notion that the phonosemantic mappings of ideophones should be “rooted in its articulation,” following previous studies (Diffloth 1994; Oda 2000; Strickland et al. 2017; Taitz et al. 2018).

Figure 1 illustrates how Hamano (1998: 40) assigned phonosemantic mappings to the CVCV root structure of Japanese ideophones. The tier structure (upper box of Figure 1) illustrates the broader categories of meaning in the CVCV context, i.e., if a phoneme is in X position it depicts a Y kind of percept (Talmy 2000). The lower box of Figure 1 shows what each phoneme specifically means in the Japanese ideophone poka-poka ‘a dull, hollow sound.’

Although Hamano’s (1998) analysis of Japanese ideophones draws conclusions which have since been disputed by some (Haiman 2018: 121–122) and revised by others (Akita et al. 2013; Nasu 2015), the basic principle remains the same for phonosemantic analyses from other languages: each phoneme depicts an iconic percept. Hamano’s (1998) tier-based analysis, though exemplary, is designed with the strict CV phonotactics of Japanese in mind and is not applicable to other languages.^[2] While intuitive yet cursory phonosemantic analyses are found throughout the language-specific chapters of Sound Symbolism (Hinton et al. 1994) and Ideophones (Voeltz and Kilian-Hatz 2001), these often verge on the impressionistic and suffer from a lack of cross-linguistic comparison. This issue can be mitigated by investigating cross-linguistic ideophone systems for their mappings between phonology and semantics. This is the route we intend to take in this study.

3.1 Database

Though language-particular databases for ideophones are becoming more widespread, e.g., the Chinese Ideophone Database (Van Hoey and Thompson 2020), the Quechua Real Words project (Nuckolls et al. 2017) or the Multimedia Encyclopedia of Japanese Mimetics (Akita 2016), there is currently no cross-linguistic database dedicated solely to ideophone inventories. We created a database of 13 languages which were selected with the aim of being as typologically diverse as possible (see Figure 1) despite the limited number of linguistic descriptions for ideophone inventories in the world. Our major criterion for selecting languages is that 40 or more ideophones were reported per source. Number of ideophones per language and language family are reported in Table 1. Due to their depictive nature, and the various methods of collection (fieldwork elicitation, dictionaries), the ideophone inventory numbers reported in Table 1 are not absolute, but instead reflect a general picture about the semantic “visibility” of ideophones per language. This is in line with a claim recently put forth by Dingemanse (2019) that ideophones form an open class, speaking to the creative potential of speakers to coin new ideophones. The languages in our database are as follows: Manyika Shona (Franck 2014), Uyghur (Wang and Tang 2014), Manchu (Xiao 2015), Chaoyang Southern Min (Zhang 2016), Ma’ai Zhuang (in prep),^[3] Kam Dong (Gerner 2005), Akan (Ofori 2009), Kisi (Childs 1988), Kuhane (Mathangwane and Ndana 2014), Pastaza Quichua (Nuckolls et al. 2017), Upper Necaxa Totonac (Beck 2008), Temne (Kanu 2008), and Yakkha (Schakow 2016). They are alternatively presented in Figure 2 according to geographic distribution.

Figure 2:

Geographic representation of the languages in the database.

3.2 Semantic features

Definitions per ideophone were entered into the database according to how they were described in the source documentation. Minor stylistic changes in the wording of some definitions were made to synthesize them across languages, e.g., “sound of a dog’s bark” and “sound of barking dog” were entered as “sound of barking dog” for consistency. These subtle differences in the word choice of the source documentation were interpreted as a product of English syntax rather than of ideophone meaning itself. If an ideophone was reported with multiple definitions, e.g., “sound of flowing water; peeing”, each was entered separately into the database, i.e., once for “sound of flowing water”, and once for “peeing”, in line with strategies set forth in the Cross-Linguistic Data Format paradigm (Forkel et al. 2018).

Definitions for reduplicated forms were entered into the database only if they were described differently from non-reduplicated equivalents. Each definition was coded with semantic features created in correspondence with Dingemanse’s (2012: 663) implicational hierarchy of ideophones (see Akita 2009; McLean 2020; Van Hoey in print for alternative approaches). Dingemanse’s (2012) hierarchy begins with monomodal depiction of sound as its most fundamental category and goes on to include four other cross-modal semantic categories: sound < movement < visual patterns < other sensory perceptions < cognitive states. For each of these categories, nine binary semantic features (Table 2) were created based on cross-linguistic ideophone research on the observations of what ideophones depict across languages (Hamano 1998; Hinton et al. 1994; Nuckolls et al. 2016; Van Hoey 2018). It is important to note that these semantic features are not mutually exclusive. An ideophone may be coded for multiple, seeing as most ideophones are multisensory (McLean 2020; Nuckolls 2019). For example, the Chaoyang /hu.hu/ ‘wind blowing’ was coded with [+sound] (because this ideophone depicts an auditory percept), [−telic] (because this ideophone does not involve a perceived endpoint of an event), [+wind] (because this ideophone involves a percept created by the movement of air), and [−motion] (because this ideophone is not depictive of a motion plus a resulting state or manner).

While the assignment of semantic features based solely on textual documentation is a far from perfect methodology when it comes to capturing the subtle nuances, multisensory percepts, and contextual meaning variations of ideophones, it is not yet clear what sort of methodology, and what extent of native speaker input, would even render such a goal possible. Hand gestures have been shown to be an insightful tool when it comes to eliciting semantic percepts of ideophones which native speakers may find too subtle to verbalize (Dingemanse 2015). There also are at least three ideophone dictionaries which make use of visual information to explain the meanings of ideophones (Akita 2016; Gomi 1989; Nuckolls et al. 2017). However, since this calibre of detailed documentation is only available for two languages so far, we are forced to contend with textual definitions for (1) determining semantic features, and (2) sentence examples (if provided) for any basis of contextual meaning variation. Therefore, our assignment of semantic features is based on inference of semantic properties which are inherent to or implied by the definitions provided in their documentation. Examples from Kuhane are given in Table 3.

Crucially, if a percept was not explicitly stated then it was not reflected in our assignment of semantic features. For example, one could imagine that the Kuhane ideophone /gwa/ ‘sound of entering abruptly’ depicts kind of visual information, rendering it [+appearance]. However, because Mathangwane and Ndana (2014) did not specify anything in their definition of /gwa/ about visual information, our semantic coding was thus ‘sound’ = [+sound], ‘of entering’ [+motion], and ‘abruptly’ [+telic]. Likewise, /tʃevutʃevu/ ‘looking around continuously’ was coded as ‘looking’ [+appearance], ‘looking around’ [+motion], and ‘continuously’ [−telic]. A more challenging example is /tʃootʃoo/ ‘whispering,’ this was coded as [+human] since it is an action done by humans, [+sound] because it is an auditory percept, and [+wind] because it involves a depiction of (sibilant) air movement. However, it was not obvious whether to code ‘whispering’ as [+motion] or [−motion]. Is ‘whispering’ an active and discriminable form of movement, comparable to that of ‘chopping’ or ‘splashing’ or is ‘whispering’ simply an auditory percept? If a semantic feature was called into question, we chose to err on the side of caution and thus refrained from coding that feature, i.e., the feature value assigned was negative by default.

A native English speaker coded the entire database and two other native English speakers who did not know the purpose of the study, checked whether they (a) fully agree, (b) maybe agree, or (c) disagree with the coding. The agreement rate between the three raters (the first author and the two independent raters) was 0.73, and Gwet’s AC1 was 0.80, which indicates quite high reliability.^[4] We filtered out 14 items out of 1,874 ideophones, where both raters disagreed with the assigned features. We would like to restate that the purpose of this study is not to provide a detailed semantic analysis of the remaining 1,860 ideophones. Rather, what we strive for is to determine whether general properties of ideophone meanings pattern according to articulatory properties of phonemes.

3.3 Articulatory features

All consonants were coded using seven binary features, listed in Table 4, according to how lips, tongue, and airflow, are involved in their articulation. Our coding follows a linear order of phonemes. Just as the semantic features are based on cross-linguistic observations of ideophones, our articulatory features are also empirically driven by previous ideophone research in that they have been shown to create lexically contrastive meanings for ideophone inventories across different languages (Diffloth 1972, 1979; Hamano 1998, 2019; Li 2007; Oswalt 1994; Strickland et al. 2017; Thompson and Do 2019). It is important to note the articulatory features here are different from traditional phonological features, like those of Chomsky and Halle’s SPE (1968) or Clements’ (1985) feature geometry. Our features illustrate contrastive movements required for a phoneme to be realized but not for a phoneme to be differentiated from other phonemes per se. For this reason, phonemes like /d/ and /l/ may be assigned an identical set of feature values. Our features referring to oral contact, tongue resting, and tongue root were designed to account for manner without reference to place of articulation. The reason for having [±tongue resting] as a feature, as opposed to [±tongue movement], was to acknowledge that tongue body and tongue tip movement may occur in some articulations but not as an active or direct result of articulation. In these cases, the tongue assumes an inactive position which may vary slightly according to the sound being made (Gick et al. 2004). See the accompanying OSF repository for a full list of phonemes coded with their articulatory features.^[5]

The binary nature of our seven features means 14 possible feature values overall. If properties of iconicity are truly universal, then we predict that the universally accessible properties captured by our articulatory features should bear the explanatory power for what perceptuo-motor affordances underpin iconicity and its notions of (analogical) depiction. While some feature values can subsume others, i.e., [±oral contact] subsumes [±labial], the decision to test the subsumable [±labial] is again to do with lexical contrasts observed. For example, in Chaoyang we have [+labial] [+oral contact] /pu.pu/ meaning ‘rapid movement’ and [−labial] [+oral contact] /tsu.tsu/ ‘whispering.’ Likewise, in Pastaza Quichua we have [+labial] [+oral contact] /pɑw/ ‘manner of being turned downward’ and [−labial] [+oral contact] /kɑw/ ‘sound of stepping on dry leaves.’ We also include the subsumable feature [+tongue root], again, given its ability to create lexical contrasts. For example, in Akan Twi we have [+tongue root] [+oral contact] /kuu/ ‘call of a large bird’ and [−tongue root] [+oral contact] /tuu/ ‘manner of hitting with the fist.’ The reason we created these subsuming features was so that the general manner of the consonant is accounted for regardless of its place of articulation in the oral tract.

The vowels attested in ideophones were coded using five binary features, listed in Table 5, according to the position of the tongue relative to the extremities of the oral cavity and whether lip rounding is involved in articulation.

5.1 Collostructional methodology

As outlined above, the goal of this study is to check if meaningful relations, such as the four predictions made above, exist between articulatory and semantic features, and why this is so. We operationalize this by using the collostructional framework (Stefanowitsch and Gries 2003; Gries 2019), in order to investigate the correlations between form (articulatory features) and meaning (semantic features). The fundamental idea of collostructional approaches consists of tallying co-occurrences of features to be investigated and placing them in a contingency table. Let us first illustrate this framework with an application from construction grammar.

Suppose we wanted to know which noun is most likely to occur in the English construction [N waiting to happen] (Stefanowitsch and Gries 2003). One could start by collecting corpus data to gather all token frequencies of nouns in this position, which would provide some measure of insight into the usage. However, it is more informative to find out that there is some degree of special attraction between the N and the constructional slot in [N waiting to happen]. This requires a method to inspect the relative strength between N and the constructional slot. One then can proceed to make contingency tables for all nouns that are identified in that construction.

The nouns found for this construction by Stefanowitsch and Gries (2003), together with their token frequency in brackets are: accident (14), disaster (12), welcome (1), earthquake (1), invasion (1), recovery (1), revolution (1), crisis (1), dream (1), it (sex) (1), and event (1). As shown in Table 6, this means that accident occurs 14 times in this construction (cell a), while 21 times there is another word that occurs in it (cell b). The token accident occurs 8,606 times in other constructions (cell c), while logically there are 10,197,659 constructional contexts that do not feature accident (cell d).

The next step in the collostructional approach is to use an association measure to calculate the relative strength between the Ns and the construction. While Gries and Stefanowitsch initially adopted the Fischer-Yates Exact test (Gries and Stefanowitsch 2004a; Gries and Stefanowitsch 2004b; Stefanowitsch and Gries 2003), which calculates the mutual strength between N and the constructional slot, there has been a shift towards directional association measures in more recent work (see Gries 2019). For example, “according” in according to attracts “to” more than “to” attracts “according”, since to is a simple preposition. Or the other way around, “instance” attracts “for” more in for instance than “for” attracts “instance”. Of course, (near) perfect attraction can exist with unique combinations, such as bona fide (Gries 2019: 393). A well-established unidirectional association measures that takes contingency into account (Ellis and Ferreira-Junior 2009; Levshina 2015) is Δ P , which comes in two variants: Δ P noun → construction and Δ P construction  →  noun , the details for calculation will be given below (see 5–6 below). This way it becomes possible to see how much a construction attracts a given N to its slot, but also conversely how much a given N attracts (or repels) that construction.

Gries (2019) suggests that what then remains is the discussion of the results after bringing together a number of indicators, such as showing both ΔPs , token frequency size, dispersion etc. In our application of this method, we are dealing with type frequencies of ideophone inventories. Consequently, we will not go as far, but will provide an extra step of linear regression on which to base our discussion with regards to the four predictions made above.

5.2 Database analysis: consonants

The application of collostructional methods to ideophone-related studies is not new, see Smith (2015) for a study on phonesthemes in Old Chinese reduplicatives, or Van Hoey (2020) for applications to ideophones as they occur in Mandarin constructions. This study adopts collostructional method for investigating associations between articulatory features and semantic features. As a reminder, the articulatory features (n = 7 × binary distinction = 14) are provided in Table 4, and the semantic features (n = 9 × binary distinction = 18) in Table 3. We counted each logically possible combination of semantic and articulatory features per language, shown for [wind] and [airflow] in Akan Twi in Table 7 as an example. Calculating this combination for other languages, together with their respective Δ P values, results in Table 8. Note that both Δ P values are calculated as follows (5–6). In these formulas, a, b, c, and d stand for cells in the contingency table, as shown in Table 7.

Based on the counts in Table 7, the Δ P of [+wind] to be realized with [+airflow] is Δ P + semantic → + articulatory = – 1 30 − 31 244 = 0.24 , and Δ P of [+airflow] to get involved in the meaning of [+wind] is Δ P + articulatory → + semantic = – 1 42 − 19 232 = 0.18 . These numbers indicate that in both directions there is an attraction between the semantic feature of [+wind] and the articulatory feature [+airflow] in Akan Twi, with the semantic → articulatory relation stronger than the opposite. In order to check if there are any associations between an active articulator [+articulatory feature] and the absence of a semantic feature [−semantic feature], e.g., the pair [+airflow] and [−wind], we have adapted the formulas for the calculation of the respective Δ P values (7–8). In practice, these values will result in the opposite polarity of the [+articulatory] ∼ [+semantic] pairs. In the case of [+airflow] and [−wind], Δ P − semantic → +  articulatory = − 0.24 and Δ P + articulatory → −  semantic = − 0.18 . We analysed the relation between [+articulatory] ∼ [±semantic], excluding the [−articulatory] ∼ [±semantic] pairs, because a core question here is how active articulatory gestures, realized as positive articulatory features, are correlated to semantic features.

The Δ P values for the combination [+wind] and [+airflow] across languages are provided in Table 8 as an example. It can be seen that different languages display different relational strengths for this semantic-articulatory pair. For this specific relation between [+wind] and [+airflow], Δ P + semantic → + articulatory overall shows higher values than Δ P + articulatory → + semantic , suggesting that it is more likely to predict an articulatory feature [+airflow] from a semantic feature [+wind] than vice versa.

It is easiest to first glance at the different combinations of Δ P statistics by visualizing them. The figures for all correlations (see Appendix) show the Δ P correlations with the nine semantic features for each of the seven articulatory features. The data points are the 13 different languages. Warm colours represent the [+articulatory] ∼ [+semantic] pairs; cool colours the [+articulatory] ∼ [−semantic] pairs. If a datapoint is situated in the upper right quadrant, it means there is a mutual attraction between semantic and articulatory feature, although not necessarily of the same strength. In the lower left quadrant, it indicates mutual repellence. Even though other relations did not occur in our data, a datapoint in the upper left quadrant would indicate that a semantic feature is more likely to attract an articulatory feature under consideration than vice versa; and a datapoint in the lower right quadrant would show that an articulatory feature attracts a semantic feature but this semantic feature does not rely on this articulatory feature at all. We have also added polygons (brown and turquoise respectively), which display the spread of the different points, as well as a linear regression line (respectively, red and steel blue) for each plot, which will be treated below, as the second type of finding. A relatively widespread polygons indicate that the strengths of the correlations across languages are disperse; a narrowly scoped area, on the other hand, indicates that languages pattern more closely together in terms of the Δ P correlations. Let us illustrate the findings with the pairs involving [+airflow] as an example in Figure 3.

Figure 3:

Δ P semantic →  articulatory and Δ P articulatory →  semantic correlations for the articulatory feature [+airflow] and the nine semantic features. Datapoints in the upper right quadrant for each panel are said to mutually attract each other; in the lower left quadrant they mutually repel each other. The polygons indicate the spread of the datapoints. Linear regression lines have been added as well. Warm colours indicate the values for [+semantic] and [+articulatory] pairs; cool colours for [−semantic] and [+articulatory] pairs.

Figure 3 shows the correlations between the articulatory feature [+airflow] and the nine semantic features under our consideration. Prediction (1a) states that [+airflow] will be associated with wind or friction. We see that all data points for the pair [+airflow] and [+wind] (Figure 3, lower right panel) are in the upper right quadrant, indicating that there is attraction from articulatory feature to semantic feature (x-axis) and also vice versa from semantic feature to articulatory feature (y-axis). It can thus be said that the prediction is corroborated. For [+airflow] and [−wind] we find the reverse: the active articulator [+airflow] is mutually repellent with regards to the absence of [wind]. Turning to the pair [+airflow] and [+friction], prediction (1b), (Figure 3, upper right panel), the story is largely the same: almost all of the language datapoints are situated in the upper right quadrant. As a consequence of these two pairs, when encountering fricatives in an ideophone, there is a reasonable probability that wind or friction is depicted. However, do note that wind and friction rely more on the articulatory features than that these features attract them. In other words, there is mutual attraction, but it is not of the same strength.

In Tables 9, 10, we show the semantic-articulatory pairs for which 11 or more of the 13 languages show mutual attraction and mutual repellence. The plausible motivations for these highly correlating pairings will be discussed below.

Let us investigate for which pairs the correlations between Δ P ± semantic → + articulatory and Δ P + articulatory → ± semantic values are very tight. We do this by calculating linear regression models for all pairs. Because we are interested in the predictive ability of the models rather than the intercept and slope values, we first inspected the F-ratio, omitting ratios smaller than 1. In this step, no models were omitted. Next, we took out the accompanying p-values that were greater than 0.05, leaving 60 pairs. Finally, we inspected the adjusted R ² for the models. Since we wanted to focus on the tightest fits, we chose an arbitrary cut-off point of 0.90. This resulted in 11 pairs. Note that this does not mean that other pairs did not have any correlation; rather, the predictive correlation between Δ P + semantic  → + articulatory and Δ P + articulatory → + semantic is the strongest for the remaining pairs, which are listed in Table 11. After the consideration of vowels (Section 5.3), the following discussion session (Section 6) will explain implications of significant correlations found from consonants and vowels against our predictions.

To sum up, both types of findings show that indeed some articulatory properties of consonants pattern to semantic features corresponding to aspects of ideophone meaning. This implies that articulatory schematic properties of phonemes, universally accessible to all speakers, are important in forming the perceptuo-motor analogies that make ideophone meaning iconic.

Below, we will discuss what may be the reason why some semantic and articulatory feature pairs display mutual attraction (Table 9), mutual repellence (Table 10) or displays a tight correlation between its Δ P values (Table 11).

5.3 Database analysis: vowels

The vowels were analysed following the same method as for the consonants. We had three main features, each with binary distinction, resulting in six features: [+front]/[+back], [+high]/[+low], [+rounded]/[+unrounded], that were paired with the nine semantic features. Like with the consonants, we have two types of findings.

As with the consonants, the first analysis concerns the mutual attraction and repellence of vowel features and semantic features. As can be seen from Tables 12, 13, there was no pair that occurred for all 13 languages in our sample. To be consistent with the consonant analysis, we set a threshold to 11 languages: if 11 or more languages show the correlation between the two features, we analysed those correlations. Under this set of criteria, we found three significant correlations (Table 12). For the pair [+motion] and [+unround], while there was an overall significant correlation between the two, most values are not clustered around the upper right edge or lower left edge (Figure 4, panel 3). This suggest that a general correlation can be found, but it is on the weaker end (the maximum absolute value is ca. |0.2|). The pair [+sound] and [+back] (Figure 4 panel 1) has some datapoints that are more skewed toward an upper right edge, indicating a somewhat stronger correlation between the two, although the range is also much wider. The pair [+wind] and [+low] (Figure 4, panel 2) showed the strongest negative values from among 11 languages. For this correlation one can putatively suggest that the sound of wind typically is perceived as and thus depicted as higher pitched, resulting in an avoidance of low vowels across ideophone systems.

Figure 4:

Δ P + semantic → + articulatory and Δ P + articulatory  → + semantic correlations for the three vowel features that display mutual repellence and mutual attraction. Note that we have used a different scale to present these three tables than the one used in all other figures.

Like for consonants, the second analysis focuses on the highest F-ratios, the concomitant p-value (<0.05) and the adjusted R ² (>0.90). As can be seen in Table 14, we find 17 pairs for which the Δ P values are quite similar: if there is a positive attraction from articulatory feature to semantic feature, e.g., [+low] and [+telic], there is also an almost equal positive attraction from semantic feature to articulatory feature. These correlations are meaningful but the distributions of the correlation points from different languages are scattered as well as clustered together, meaning that there is no clear systematic patterns in terms of the directions of the correlations. This distributional tendency is different from the analysis of consonants, where almost all correlations were found from similar clusters, indicating that languages show similar distributions of each correlation (Table 15).