Iconic features in nominal reduplication

2.1 Plurality, collectivity, and distributivity

In my data set, reduplication most frequently marks some form of plurality, confirming results from previous work (Schwaiger 2017:85; Mattiola and Barotto 2023:156). A typical example of plurality marked by reduplication is (1), repeated here; another example is provided in (4):

The pattern in (1) is only one of many ways to express plurality in Hausa. Elsewhere in Hausa, this pattern has a distributive meaning (discussed below), but in these cases, it expresses an unmarked plural meaning (Newman 2000:457): it conveys that its referent consists of multiple entities, which can each be referred to with jōjì, but it does not tell us anything more than that. The examples in (4) are similar. We can define plural as follows:

There are two subtypes of plural. First consider Nakanai: in this language, nouns are not normally inflected for number (Johnston 1980:167–168); instead, a plural pronoun can be used if plurality needs to be expressed (6a). However, to refer to a set of entities that is conceived of as a single group, the reduplicative pattern in (6b) can be used:

Thus, the reduplicative pattern in (6b) can be contrasted with the simple non-collective plurals in (1,4). The Nakanai construction in (6b) is not only marked for plural, but also for collective:

The case of Chimakum is slightly different again. In this language, “[t]he plural is, more properly speaking, a collective, but is frequently used in a way similar to our plural, namely, when the collective and plural ideas nearly coincide” (Boas 1892:39). Though Boas does not give examples in context, I take this to mean that the reduplicated forms in (8) can refer to collectives when such an interpretation is promoted in the context, but that the forms can also be used in non-collective contexts.

This is different from the Hausa pattern in (1), which cannot be used to promote a collective reading, but it is also different from the Nakanai pattern in (6b), which is infelicitous outside collective contexts. Instead, the Chimakum pattern in (8) has a plural-collective interpretation as well as a plural interpretation. Such a pattern thus receives two entries in my data set.

There are various phenomena apart from reduplication that point to grammar being sensitive to a collective feature, and thus justify its inclusion in my analysis. For instance, in English, nouns referring to collectives can often trigger unexpected plural agreement (e.g., committee; Corbett 2000:188–191), and certain verbs assume collectivity of their subject, as in The ten girls gathered as opposed to The ten girls smiled (Champollion 2020).

Frequently discussed in relation to collectivity is the notion of distributivity. This notion can be illustrated on the basis of Cahuilla:

Here, the reduplication of ki ‘house’ indicates simple plurality, but also expresses that the houses are distributed in space (the implication that the houses are usually empty seems to be due to partial lexicalization, as it does not appear in related languages). This pattern is marked for plural-distributive:

The “abstract space” in this definition refers to abstract domains which can be represented metaphorically as a space, such as the domain of time or color. For the latter, for instance, visible colors can be represented as points in a three-dimensional abstract space with axes for light intensity on the three wavelengths recognized by our cones (Churchland 1986:299–302). The following is an example of a distributive pattern referring to an abstract space:

Rooryck (2024:20) draws attention to a set of quantifiers in English that “insist on the variety and differentiation among the units of the plural set”: various; different; sundry; miscellaneous; diverse; all manner of. This can be seen as evidence beyond reduplication for the existence of a distributive feature.

By definition, both distributive and collective imply plural. One might imagine cases where distributive and collective co-occur (e.g., to refer to a group of sheep scattered in a field), but I did not find such cases in my sample. It seems that by considering entities as part of a group, we conceive of them as close to each other and therefore not distributed over the salient abstract space based on group membership.

In some languages, the plural does not seem to be limited to count nouns. In Igbo, the same pattern can be used to express the plural of count nouns and a large quantity of a mass noun (12). As discussed in Section 2.3 below, there is no evidence for an ‘augmentative’ feature: the seemingly augmentative meaning in (12c–d) only appears with mass nouns, marked by a pattern that also marks plurality of count nouns. Therefore, I treat it as an extension of plural to mass nouns.

There are two reduplicative patterns in my data set that might perhaps be analyzed as marking singulative number. However, because it concerns so few examples and their function is not exactly clear, I do not define a feature for this meaning at present.^[12]

Finally, some languages have been described as marking dual number with reduplication. Two noun classes in Panamint (Uto-Aztecan; pana1305) show this behavior, although it should be noted that one of them also marks the plural by reduplication (McLaughlin 2018, p. 366). Since minor numbers do not appear as a function of reduplication in my sample, I will not discuss them further here.

2.2 Similarity

In all the examples discussed in the previous section, the reduplicated form refers to a set of entities that could each be referred to using the base form. Thus, in Hausa, jōjì∼jōjì refers to a set of judges, each of which can be referred to with jōjì ‘judge’ (see (1) above). There are also instances where this is not the case, as in (3), repeated here, and (13):

The reduplicated forms in (3) refer to multiple entities, and are thus plural. However, some of the entities denoted by the set paaNii ∼ vaaNii ‘water and the like’ cannot be referred to by paaNii ‘water’ (e.g., ‘tea’ or ‘milk’). The pattern in (3) can thus be said to be plural-similar:

I also include under the notion of similarity cases where entities associated with the base are included. Thus, the following pattern is tagged as plural-similar as well:

While the kind of plural-similar meaning in (3,13,15) is frequent, the examples in (16)–(18) show that similar can also occur without plural. These examples also show different kinds of similarity that can be expressed, such as similarity involving diminution, facsimile, or imitation (17), as well as lexicalized similar meanings (18).^[13]

2.3 Diminutive and intensive

The boundary between similarity and diminution can be blurry (cf. Jurafsky 1996:553–554). Consider the following data:

In this case, Thompson and Thompson (1971) are explicit that the first two reduplicated forms have a “specialized meaning”, whereas the third is a plain diminutive. It is not always clear to what extent the judgment that a form expresses a meaning of diminution is influenced by the availability of different terms in English. However, there are many cases of reduplicative patterns that are described as marking plain diminutive meaning, without any mention of similarity. In such cases, grammars often do not even translate the diminutive form:

Since examples like (16) do not involve diminution at all, and examples like (20) do not involve similarity at all, there is enough reason to define a diminutive feature in addition to similar:

As (22) shows, diminutive can also co-occur with plural:

As mentioned before, there is no evidence for an “augmentative” feature in my sample (cf. Mattiola and Barotto 2023:149 n. 10). This is interesting, since it seems intuitively a more iconic feature for reduplication than diminutive. Examples that do seem to involve an increase in size are, in fact, of different types. The extension of plural to mass nouns in the meaning ‘a large quantity of’ has already been mentioned in Section 2.1 in relation to a pattern from Igbo (12). The same pattern provides examples that come perhaps closest to what one might call augmentative:

However, as already seen in (12), the same pattern is used for plurality of count nouns, as well as large quantities. Furthermore, the potentially augmentative reading in (23) is only found with the verb pà ‘carry’, and all examples involve body parts. It therefore looks like the augmentative reading arises because plurality is excluded, and the body part is interpreted as a mass: it is unlikely that the subject has multiple necks, so the interpretation becomes ‘lots of neck’, i.e., ‘a very long neck’. The mass reading is reinforced by pà ‘carry’.

Other examples that seem to have an augmentative meaning are really more specific, and involve a great intensity. Though some of the relevant examples can be translated into English with ‘great’, there are no examples involving concrete entities where the reduplicated form marks a greater size:

For these meanings, the following feature is defined:

This feature can be related to degree readings of attributive size adjectives which target a salient dimension of the modified noun, as in huge goat cheese enthusiast, where huge refers to the dimension of enthusiasm rather than spatial size (see e.g. Morzycki 2009).

Although intensive meaning is very common in verbal reduplication (Lǐ and Ponsford 2018), this feature occurs with nominals in only four languages in my data set: Bikol (Austronesian, biko1240); Central Cagayan Agta (Austronesian); Luvale (Atlantic-Congo, luva1239); and Modern Greek (Indo-European, gree1276). In none of these cases does it co-occur with other features. However, due to the low frequency of this feature, this may be accidental.

2.4 Exhaustiveness

In a number of languages, reduplication expresses the notion of ‘every …’ or ‘all of …’. Consider:

This may seem to be an extreme case of distributive, in which the reduplicated form does not denote ‘entities here and there’ but ‘all entities, everywhere’. However, while distributive is clearly dependent on plural, there appear to be cases of reduplication that are arguably similar to (26)--(27) but do not involve plural: in these cases, the universal quantification does not apply to elements of the set of entities that can be denoted by the base, but to the extent of one single entity:

Though the Agta pattern in (28) is the only example of exhaustive without plural, we can tentatively define:

This notion can be linked to distributive quantifiers like each, which likewise apply universal quantification. It also corresponds to the notion of “global plural” recognized by Corbett (2000:30–33).

2.5 Exclusiveness

Finally, there are cases where reduplication expresses that a predicate from the context applies only to the entity/entities denoted by the base noun. This usually co-occurs with plural (30)–(31), but there are examples without plural (32):

We can define:

Like the other features, this is a notion that is commonly grammaticalized. For instance, Bonomi and Casalegno (1993) show that English only associates universal quantification with a domain marked by focus. In languages like Hindi, Amharic, and Indonesian Bajau, the exclusive focus that is marked by stress in English is instead marked by reduplication.

2.6 Combinations

With the eight features discussed above, the meaning(s) of all reduplicative patterns in my data set can be accounted for.^[14] The features are largely independent of each other. Almost all features can co-occur with at least plural (the exception is intensive, which occurs in only four languages). Apart from distributive and collective, all features can also occur without plural. Additionally, complex combinations occur, such as plural-collective-similar-exhaustive (e.g. Pacoh ʔa.lɔːŋ ‘tree’ → ʔa.lɔːŋ∼ʔa.lɛː ‘vegetation’; Austroasiatic, paco1243; Alves 2006:38). In total, 19 different sets of features are attested in my variety sample.

3.1 Current approaches to apparent non-/counter-iconicity

There are several ways to deal with the issue of seemingly non-iconic and counter-iconic meanings. Non-iconic meanings, which are neither iconic nor counter-iconic, are not immediately problematic as they may have developed from earlier iconic meanings through regular semantic shifts (e.g. Fischer 2011). However, counter-iconic meanings are more problematic, as we might expect them to be actively avoided; they may be difficult to understand and acquire. For this reason, most research has focused on the apparently counter-iconic diminutive.

A relatively early explanation for the diminutive meaning can be found in Regier (1994), who proposes that the meaning ‘small’ is derived from an association with baby talk. Regier observes that many languages have reduplicative lexemes for early vocabulary, such as English baby and daddy, “yielding an associational link from repetition to baby” (1994:6, emphasis original). The move to express ‘small’ then comes from “the perceptually very salient fact that babies are small” (Regier 1994:6). This explanation seems to miss a few steps, jumping from utterances by small things (baby talk) to a derivational device to express smallness (diminutive), and has not found broad acceptance (cf. e.g. Maas 2005).

Another explanation is proposed by Kouwenberg and LaCharité (2005). They suggest that the diminutive function of reduplication in Caribbean creoles may have arisen from an earlier dispersive function: something that is ‘yellow here and there’ is also ‘a little yellow’, which might explain a semantic shift from plural (and/or distributive) to diminutive. Lee (2023:15) also notes an association between plurality and diminution in reduplication in Formosan languages, but explains it from biology: during reproduction, multiple small members are created, and so plural and diminutive naturally co-occur (for further discussion of a possible link between plural and diminutive, see Schwaiger 2017:94–98). In my sample a combined plural-diminutive meaning occurs in only five languages: Central Nahuatl (Uto-Aztecan, cent2258), Halkomelem (Salishan, halk1245), Ineseño (Chumashan), Mbula (Austronesian, mbul1263), and Tillamook (Salishan, till1254). The fact that it occurs in various language families and macro-areas suggest that there is indeed some cognitive link between the two features. However, in the vast majority of languages where reduplication is used to mark diminutive there is no evidence for an additional plural meaning. It therefore does not seem likely that such a link can explain all diminutive patterns found across languages.

Other authors have sought to reduce all meanings commonly expressed by reduplication to a single iconic core. For example, Stolz proposes that “[w]hat is iconic […] is the relation that holds between the linguistic sign representing the derivational basis and the linguistic sign representing the diminutive because the latter is conceptually more complex and at the same time also receives the more complex expression” (Stolz 2007:345; cf. Abraham 2005 for a similar idea). Thus, in his view, both diminutive and plural are iconic, because both are more complex meanings than that of the non-reduplicated form. This hypothesis is problematic in that it seems to overgenerate in two ways. First, certain meanings that are cross-linguistically marked, like definiteness, are, contrary to prediction, rarely or never expressed by reduplication (for full reduplication, see Stolz et al. 2011:194, who list many more functions that are not marked by full reduplication). Second, the explanation applies to reduplication and affixation alike, since regular, non-reduplicative affixes also create more complex forms. Thus, we would expect the functions of reduplication and affixation to largely overlap, but this is not the case: reduplication is cross-linguistically used for a rather specific subset of the meanings that can be expressed through affixation. In addition to these arguments specific to reduplication, Haspelmath (2008:6–14) has argued that the idea of an “iconicity of complexity” in general makes the wrong predictions, and that the apparent examples of such an iconicity are better explained by a more general economy principle.^[15]

In a similar but slightly more specific vein, Mattes argues that “[e]very reduplicated word form which expresses any kind of quantity change with respect to the meaning of the base (i.e. intensity, plurality, diminution, etc.) is an example of ‘iconic’ reduplication, because the change of quantity in meaning corresponds to a change of quantity in form” (Mattes 2014:121). However, while this abstraction captures diminution, it does not provide a cognitive explanation why a decrease in quantity should be marked by an increase of phonological content. Similarly, Mattiola and Barotto (2023:159–163) propose specifically for nominal reduplication that all functions can be seen as a kind of “change of referents’ specificity”. This notion is intended to capture plural meanings, which decrease specificity since plurals are always to some extent generic, while also capturing the diminutive: diminutives increase the specificity as they limit the set of entities that can be denoted by a form (i.e., diminutives refer to the subset of small Xs, rather than the set of all Xs). This approach still suffers from the same issue of overgeneration as that of Stolz (2007), namely that definiteness (clearly an increase in specificity) is not among the functions of reduplication. It also suffers from the issue with Mattes’s (2014) approach that the cognitive underpinning of the proposed iconicity is not quite clear.

3.2 An alternative: fine-grained iconicities

To sum up, previous accounts of reduplicative iconicity rely either on assumed semantic shifts whose operation can be demonstrated in only a small group of languages (e.g. ‘baby’ → diminutive in Regier 1994 or plural/distributive → diminutive in Kouwenberg and LaCharité 2005; Lee 2023), or on overly general iconic principles, such as complexity (Stolz 2007) or change in quantity or specificity (Mattes 2014; Mattiola and Barotto 2023). A further problem arises because previous studies have generally lumped together the different formal types of reduplication. However, if reduplication is iconic – that is, meaning is tied to form – we might expect that full, partial, and echo reduplication express different meanings.^[16]

The idea I would like to pursue here is that all of the features described in Section 2 are iconic, but that the iconicity of each semantic feature is based on a different formal property. The fine-grained nature of this approach avoids sweeping generalizations to safeguard against overgeneralization. Furthermore, making the formal aspect motivating each feature explicit renders the hypothesized iconicity falsifiable using data of the type collected in Section 2.^[17]

My approach is based on that of Lǐ and Ponsford (2018:61, 77), who distinguish five different possible iconicities – “iconic relationships between specific aspects of form and meaning” – for verbal reduplication.^[18] In their analysis, functions can be iconic in these five different ways, and different iconicities can be exploited at the same time. For instance, “habitual” is motivated by an iconicity of Identity (identical form reflects identical events) as well as Magnitude (larger form reflects multiple events). I adopt this approach but deviate in two ways. First, whereas Lǐ and Ponsford describe functions as being iconic in different ways, I link iconicities directly to features. The iconic expression of complex functions falls out of their reduction to sets of features. Second, whereas Lǐ and Ponsford limited themselves to full reduplication, I include partial and echo reduplication as well. Lǐ and Ponsford’s reason to exclude these formal types was that in full reduplication it is “clearer that the doubling of the base is noticeable to speakers and therefore in principle manipulable by them for expressive purposes” (2018:78). This may be true for doubling, but I also explore iconicities that are not based on doubling but on formal aspects specific to partial and/or echo reduplication. Thus, in my approach, full and echo reduplication alike can iconically express plural (because the form, like the denoted entity, contains multiple identical elements), but only echo reduplication can iconically express similar (because only in echo reduplication is the formal copy different from the base, like the denoted entity is different from that denoted by the base). Therefore, echo reduplication is predicted to be more likely to mark plural-similar, whereas other formal types are predicted to mark other functions involving plural more often.

It is important to stress once again that the iconicities described in this section are hypotheses that will still be tested against the data in Section 4. For an overview of the predictions made by the iconicities described here, see Table 1 in Section 4.2; for an overview of the results, see Table 4 in Section 4.3. In what follows I provide the rationale for each of the hypothesized iconicities.

Among the iconicities recognized by Lǐ and Ponsford (2018), there are two that are commonly proposed in the literature:^[19]

Based on Magnitude we can see both plural and intensive as iconic features. On the other hand, Identity cannot be linked to any features, but could be said to be linked to meanings without a similar feature. Focusing on the formal side, both Magnitude and Identity apply to all formal types of reduplication: in each type there is some identical content that is represented multiple times (namely, in the base and in the copy). As a result, these iconicities do not generate any predictions regarding the distribution of meaning over different formal types, and I will not be able to test them here.

However, by carefully looking at the formal differences between different types of reduplication, we can hypothesize other iconicities which do distinguish between formal types. In the following subsections I introduce four such possible iconicities, without any claim to exhaustiveness and recognizing that three out of four have already been proposed in one way or another in the literature. The value in revisiting them is that they will be tested on a typological sample for the first time here (see Section 4). Together with Magnitude, these iconicities can account for all eight features described in Section 2 (though, as we will see in Section 4, not all iconicities can be confirmed by my statistical tests).

First of all, echo reduplication is unique in phonologically modifying the base (as opposed to merely removing or adding phonological material), thus creating a copy which is similar to but different from it. We may expect this phonological similarity to iconically mark semantic similarity:

In terms of features this relates to similar. It is therefore predicted that similar meanings are more likely to be expressed by echo reduplication than by other types of reduplication, which do not involve phonological distortion. The possibility of such an iconicity was already recognized by Stolz (2008:127) and Kallergi and Konstantinidou (2018:93), and it is investigated in more detail by Griffin (In prep.).^[23]

On the other hand, partial reduplication is unique in creating copies that are smaller than the base. Following the same idea, we may expect this to iconically mark semantic smallness (cf. Mayerthaler 1977:28):

This iconicity clearly applies to diminutive. Therefore, based on this iconicity we predict that the expression of diminutive is more likely with partial reduplication.

A reviewer suggests that according to Smallness, diminutive ought to be expressed by subtractive morphology: partial reduplication still makes the form larger, so how can it iconically mark diminution? It is true that subtraction seems well-suited to express diminution, but this does not explain the frequent use of reduplication to mark diminutive. It may be relevant that subtractive morphology is rarely productive: if this is in part because it introduces ambiguity (mapping multiple forms onto the same form), partial reduplication could be seen as a way to use subtraction (in the copy) without introducing ambiguity (by keeping the base). Since it includes a smaller form, it can iconically mark diminutive, even though the resulting form is larger.

What sets apart full reduplication from both partial and echo reduplication is that it leaves the base intact. This is more marked than the formal aspect of the Identity iconicity defined in (35) above, for which any identical element between the base and the copy is sufficient: full reduplication is the only type in which all of the phonological features of the base are copied. I hypothesize that the fact that all phonological features are copied may iconically reflect application to all elements in a set – that is, universal quantification:

In terms of features from Section 2, this iconicity applies to exhaustive and exclusive. According to Completeness, exhaustive may be iconic because it expresses universal quantification over the entities that can be denoted by the base. Furthermore, exclusive may be iconic because it expresses universal quantification over the set to which a predicate from the context applies: the meaning “he invited the students only” (from (31) above) can be rephrased as “all the ones that he invited were the students”.^[24] Thus, the prediction is that exhaustive and exclusive are more likely to be expressed by full reduplication.

A final formal difference between the different types of reduplication is the tightness of the bond between base and copy. Stolz et al. (2011:48–52), who focus on full reduplication, describe instances of reduplication as “non-contiguous” when base and copy are not immediate neighbors (cf. also the different types of repetition described by Thornton 2023). When partial reduplication is also considered, it becomes clear that there are differences in the degree of contiguity even between patterns where base and copy are immediate neighbors. Concretely, partial reduplication always creates forms that are a single phonological word, whereas full and echo reduplication may create forms that consist of two phonological words. More generally, the phonological bond between base and copy is at least as strong in a partial reduplication pattern as it is in a full reduplication pattern. We may expect this greater phonological independence of base and copy in full and echo reduplication to iconically mark entities that are clearly separate. This brings us to the following hypothesized iconicity, adopted from Lǐ and Ponsford (2018):

On the formal side, this iconicity applies to full and echo reduplication, where the base and the copy are relatively discrete.^[25] In terms of semantics, Discreteness can be linked to distributive, which expresses that different entities are discretely distributed in space. At the same time, non-discrete forms can be linked to collective, which expresses that entities are seen as part of a group and therefore less discrete. Thus, the prediction is that distributive is more likely with full and echo reduplication, whereas collective is more frequent with partial reduplication.

4.1 Data set

In this section the predictions made in Section 3 are subjected to statistical tests. For this purpose, the unbalanced variety sample from Section 2 was filtered to create a probability sample balanced for areal or genealogical affiliation, following a top-down procedure similar to the one advocated by Bickel (2008). Bickel suggests considering, for a given level in the genealogical tree, whether the distribution of the values of the languages in that subtree is likely skewed due to genetic relatedness. If so, all but one of the values of the majority value are removed from the data set, to avoid oversampling due to shared retentions. In this paper, the variables of interest (morphological type and semantic features) pertain to individual reduplicative patterns and not to languages as a whole (contrary to a variable like, say, word order). For this reason, the sample is a sample of patterns rather than languages.^[26]

Although Bickel (2008) provides an algorithm to apply the procedure he proposes, I followed this procedure manually, making a subjective judgment at each node in the genealogical tree. There are several reasons for this. First of all, some of the features are related. For instance, partial reduplication is used in every Salishan language in the data set to mark plural, and frequently has a distributive and/or collective value in this family as well. These latter features do not appear often enough to be able to trace them back to the root node (Proto-Salishan) based on a basic statistical test on the raw numbers. However, given that plural marking occurs in every language, and that collective and distributive are closely related to plurality but are also sometimes overlooked by grammarians, it is safe to assume that the high frequency of collective and distributive functions in Salishan is also due to genetic relatedness. In such cases I have taken the liberty to use only one entry to represent the collective/distributive function. A second reason is that in some cases patterns from related languages may have the same formal and semantic features, but it is unlikely that this is due to genetic relatedness. For example, two patterns of partial reduplication may be clearly distinct morphologically, or the specific type of similarity may differ. In such cases I have decided against lumping together the patterns.

Finally, patterns from different macro-areas were never lumped together, and I have not attempted to filter out patterns that have been borrowed through language contact from the probability sample. In particular for full and echo reduplication, it has been argued that these patterns are more susceptible to borrowing precisely because of their iconic status (Klamer and Saad 2020:316–317; Stolz 2008:127). Thus, borrowed instances are useful data points to determine the degree of iconicity of a reduplicative pattern.

After filtering, the probability sample consists of 134 reduplicative patterns from 118 languages, and 97 languages without nominal reduplication (these patterns are marked in the Supplementary Material). Figure 1 presents the geographic distribution of formal types of reduplication in the probability sample (high resolution versions of the maps in this article can be found in the Supplementary Material). Note that this representation does not distinguish different morphological types of echo and partial reduplication (for instance, although most patterns of echo reduplication in Eurasia use onset replacement, the consonant with which the onset is replaced varies). The distribution of both partial and echo reduplication over macro-areas is statistically significant (p = 0.001).^[27] Post hoc comparisons between pairs of macro-areas indicate that Eurasia deviates in both cases: languages in this area display significantly more echo reduplication than all other macro-areas (0.000 < p ≤ 0.022), and significantly less partial reduplication than Australia, North America, and Oceania (0.000 < p ≤ 0.005).^[28] For echo reduplication, Stolz (2008) has argued that its high frequency in Eurasia is due to language contact (see also Southern 2005). I am not aware of an explanation for the low frequency of partial reduplication in this area. The distribution of full reduplication is significant as well (p = 0.002), but in this case pairwise tests indicate that there only is a significant difference between North America and Oceania (p < 0.001).

Figure 1:

Areal distribution of formal types of reduplication.

4.2 Methods

The various iconicities, with their affinity with different formal types of reduplication as well as semantic features, are listed in Table 1. Of course, it should be kept in mind that iconic patterns are not immune to normal diachronic developments, including semantic shifts (cf. Fischer 2011). This may cause some reduplicative patterns to be used for functions that cannot be seen as iconic, i.e., for which there is no cognitive reason why reduplication should be well-suited to mark them. Thus, there is no need to explain all form-meaning pairings as iconic, only to identify tendencies in the data. Based on the associations in Table 1, we predict the following:^[29]

We can now determine the correlation of each semantic feature with the different formal types of reduplication in the probability sample. Consider prediction (40c) that similar is relatively likely to be expressed by echo reduplication. The relevant contingency table, in which each reduplicative pattern in the data set is one data point, is shown in Table 2.^[30] Each cell lists the number of data points of a certain formal type with/without the similar feature and, in parentheses, the number of data points expected if there were no correlation between formal type and the feature.^[31]

Whether the observed values are significantly different from those expected under the null hypothesis can be tested with Fisher’s exact test. First, a one-tailed test is used to compare the distribution of echo reduplication to the distribution of the other two types together, to evaluate the overall effect of patterns with versus without phonological distortion. For similar, a one-tailed Fisher’s exact test indicates that there is sufficient evidence to reject the null hypothesis (p = 0.003). In other words, there is sufficient evidence that echo reduplication is more likely to mark similar than other types of reduplication.

In the case of a significant result (as in the case of similar), post hoc comparisons between pairs of formal types (in this case, echo vs. full and echo vs. partial) are used to test which cells of the contingency table are responsible for the correlation. These tests also use a one-tailed Fisher’s exact test, with Holm-Bonferroni correction. For similar, the results indicate that echo reduplication is significantly more likely to express similar than both full and partial reduplication (both p = 0.009).

We can conclude that similar is not evenly distributed over formal reduplication types due to an overrepresentation of echo reduplication patterns.^[32] Furthermore, echo reduplication is significantly different from both full and partial reduplication in this regard. This confirms the prediction made by the hypothesis in (40c) and thus provides evidence for the Distortion iconicity.

4.3 Results

The same approach can be used for the other features and the corresponding predictions. All contingency tables are collected in Table 3, and Table 4 lists the results. For collective and distributive, only patterns specified for plural were counted.

There is a clear correlation of echo reduplication with similar, which persists in pairwise comparisons with both full and partial reduplication. Furthermore, exhaustive and exclusive are more likely to be expressed by full reduplication. For exhaustive, only an overall effect can be observed, whereas for exclusive, a pairwise effect when comparing to partial reduplication obtains as well.

For the significant correlations, Cramér’s V can be used as a measure of the relative strengths of the associations. The three correlations are roughly equally strong, with echo – similar and full – exclusive being the strongest (V = 0.21), followed by full – exhaustive (V = 0.16).

4.4 Correlations with plural and intensive

Recall from Section 3.2 that because Magnitude applies to all types of reduplication, the expression of plural by reduplication can be seen as iconic, but no preference for any one type of reduplication can be predicted. However, since there is no claim that the list of iconicities is comprehensive, there is no hypothesis that plural is evenly distributed either. What can be said at this point is that for the data collected here, a two-tailed Fisher’s exact test does not provide evidence for a correlation between plural and any formal type (p = 0.437). There is therefore no need for a post hoc explanation of any such correlation.

Like plural, the feature intensive is also associated with all types of reduplication through Magnitude. I suggested that it may still be less frequent with partial reduplication, because it is negatively associated with Smallness. The low number of occurrences of intensive in my sample makes it impossible to test this prediction, which is anyway relatively weak due to the interference of Magnitude. As with plural, a two-tailed Fisher’s exact test does not suggest any correlation between intensive and a specific formal type (p = 1.000).

4.5 Discussion

Based on the results from the previous subsections we can draw a number of conclusions corresponding to the predictions in (40) above. For each prediction the result is listed in Table 5.

Distortion. Three predictions corresponding to two iconicities can be confirmed straightaway. One of the strongest correlations is between similar and echo reduplication (40c). This association had already been predicted in the literature (Stolz 2008:127; Kallergi and Konstantinidou 2018:93) but was tested and confirmed with a suitably large data set for the first time here. It can be seen as evidence for the Distortion iconicity.

Figure 2 presents the geographical distribution of similar patterns. Although this meaning is attested for reduplicative patterns throughout the world, echo reduplication is rare outside Eurasia. As discussed in Section 4.1, the high frequency of echo reduplication in this region is likely due to language contact. However, not all patterns can be traced back to a single source, and in any case, the spread of a pattern through language contact can be taken as evidence for iconicity as well (on echo reduplication in particular, see Southern 2005:53). A logical conclusion is that while Distortion can ease the spread of existing echo reduplication patterns with similar meaning, echo reduplication patterns are not easily freshly coined. Thus, Distortion is a strong iconicity, but geographically restricted. The fact that similar meaning is frequent with other types of reduplication outside Eurasia suggests that the expression of this function may also be motivated by other factors. More work is needed to clarify the ways in which other patterns with similar meaning arise.

Figure 2:

Areal distribution of formal types of reduplication marking similar.

Completeness. There is also sufficient evidence for the Completeness iconicity, according to which the complete copying in full reduplication can reflect universal quantification in meaning. The two predictions following from this iconicity ((40f) and (40g)) were both confirmed. These meanings are not unique to any one macro-area (as can be seen from the maps in the Supplementary Material).

Smallness. A correlation between diminutive and partial reduplication, often suggested in the literature (Mayerthaler 1977:28; Mattiola and Barotto 2023:157), could not be confirmed. However, it should be noted that the distribution in the contingency table is skewed as predicted (only partial reduplication is used more for diminutive than expected under the null hypothesis), and that the p-value approaches the significance threshold. It may therefore be the case that this iconicity is weaker than the others, and could still be confirmed with a larger data set.

A reason for this iconicity to be weaker could be that some diminutive meanings derive from original plural meanings. An association between these meanings has been proposed for various languages (e.g. Kouwenberg and LaCharité 2005 on Caribbean creoles and Lee 2023 on Formosan languages). Even if such an association cannot explain all or most instances of diminutive, they may account for some cases where diminutive is expressed by non-partial reduplication. Alternatively, more formal information on the different diminutive patterns may lead to more precise results. The smaller the size of the copy, the more likely a reduplicative pattern would be to express diminutive, according to Smallness. Furthermore, as observed by Olivia Griffin (pers. comm.), partial reduplicative patterns with larger copies (e.g., entire syllables or feet) are indistinguishable from full reduplication for mono- and/or disyllabic roots, which may impair Smallness. Thus, including more precise morphological information in the data set may help to determine whether Smallness obtains.

Discreteness. It was predicted that the Discreteness iconicity would link collective to partial reduplication and distributive to full and echo reduplication. These predictions were not borne out, although the relevant contingency tables are generally skewed in the predicted direction. In the case of collective, only partial reduplication is used more to express collective than expected under the null hypothesis (as predicted); it is possible that a significant effect could be observed with a larger data set. As for distributive, full reduplication is used more than expected under the null hypothesis (as predicted), but echo reduplication less (contrary to prediction).

It is important to remember that Discreteness was based on the tightness of the phonological bond between base and copy. The distinction between full, partial, and echo reduplication is a relatively indirect measure of this variable. More precise phonological information in the data set may allow us to better measure the formal side of the proposed iconicity. Alternatively, the correlation of echo reduplication with similar may simply be stronger than that with distributive – assuming that distributive-similar is not a common enough meaning to receive dedicated morphology, echo reduplication would be less likely to express distributive, because it would be overruled by similar. Such an issue could be addressed by incorporating ideas about the relative strength of the various iconicities when setting up hypotheses. In a two-tailed post hoc test, the distribution of discreteness over formal types is significant (p = 0.011), so there is reason to continue looking into this.

Lack of correlations with partial reduplication. It is striking that neither prediction involving partial reduplication could be confirmed, though in both cases the distribution is skewed in the predicted direction. This is in contrast to full and echo reduplication, for which most predictions could be confirmed. This suggests that partial reduplication may be less susceptible to these iconicities for an independent reason. It has been suggested that partial reduplication may arise from full reduplication as the result of grammaticalization (Lehmann 2015:139; Bybee et al. 1994:166; and Stolz et al. 2011:164 see this as one of multiple ways for partial reduplication to arise). This is consistent with the fact that partial reduplication is less frequent than full reduplication in creoles (Stolz et al. 2011:173): since creoles have grammatical features that are significantly different from those of their parent languages, we expect to find a relatively large amount of relatively young morphology.^[33] Acquisition data suggest a similar story: full reduplication is easily coined by children even in meanings for which it is not part of the adult language, but partial reduplication is both more difficult to learn and more rarely coined in new meanings (Stolz et al. 2011:177). If partial reduplication is indeed less easily freshly coined, it is also less likely to have an iconic motivation. For this reason, it would be worthwhile to expand the data set and/or use more precise tagging in follow-up studies, to be able to find support for possible weaker effects.

5.1 An abstraction on the feature level

To highlight the benefits of an abstraction on the level of atomic features, I briefly compare the approach taken in this paper with that of Mattiola and Barotto (2023), which served as the starting point for the semantic analysis here. Mattiola and Barotto (2023) recognize 19 different functions of nominal reduplication. Some of these functions are highly specific and occur only a handful of times in their sample (only five functions occur ten times or more).^[34] By focusing on features, relations between functions become explicit (such as between plural, plural-collective, and plural-distributive, which share a plural feature). Mattiola and Barotto (2023:144–155) do recognize some clusters, including “plurality”, but functions are only ever part of one cluster and the grouping is ad hoc (for instance, “set construction”, which roughly corresponds to plural-similar, is not part of the “plurality” cluster).

The different approach has repercussions for the association of semantic properties and formal types. Mattiola and Barotto (2023:157–158) make four observations in this respect: “additive [regular] plurality” and “diminution” are marked relatively frequently by partial reduplication, and “distributive plurality” and “new referent” are marked more often by full reduplication. They do not report on associations of clusters as a whole, which are, under the view outlined in Section 3.2, more likely to be due to iconicity. Such associations could easily be investigated in this paper, where features provide a more rigorous alternative to clusters. Concretely, I found no evidence for an association of plural with any formal type of reduplication. I do see a trend according to which diminutive is more frequent with partial reduplication, but it is not significant; the same holds for the correlation between distributive and full reduplication.^[35] In sum, while an analysis on the level of complex functions remains useful for more detailed semantic descriptions, an analysis in terms of atomic features provides a sounder basis for statistical testing of iconicities, which are assumed to target a lower level of abstraction.

5.2 Iconically motivated features

Three of the features in my analysis have been shown to be iconically motivated.^[36] This claim is not merely based on intuition, as is often the case in previous studies; the evidence comes from the fact that these features show preferences to be expressed by certain types of reduplication in which the correspondence between form and function is strongest. The strongest iconicity (Distortion) links similar to echo reduplication, a type which thus marks semantic similarity through phonological distortion of the base ((36) above). This iconicity was already suggested in the literature but was tested and confirmed for the first time here. Another iconicity (Completeness) links full reduplication to features involving universal quantification (exhaustive and exclusive); this aspect of the meaning is reflected by the complete copying of the base (38).

I have not yet been able to confirm iconicities linking collective and diminutive to partial reduplication – collective because partial reduplication creates tighter phonological bonds (39) and diminutive because the copy is smaller than the base in partial reduplication (37). However, as discussed in Section 4.5, there may be confounding factors that cause the effect of iconicities targeting partial reduplication to be weaker. The relevant distributions are skewed in the predicted direction (just not significantly so), so there is reason to continue to study these possible iconicities with larger and/or more precisely tagged data sets. It would also be useful to investigate whether there is a difference in meaning between partial reduplication patterns that arise through different processes (in particular, phonological reduction from full reduplication vs. fresh iconic coinage).

Besides collecting more patterns of nominal reduplication to test for weaker iconicities, an obvious way forward would be to broaden the scope of the analysis to include other categories. Some semantic features defined here for nouns are relevant to verbs and adjectives as well. A closer look at reduplication with these categories will therefore yield not only a larger set of features, but also more data for the features identified here. In the lexical domain, iconicity is generally more common with verbs and adjectives than with nouns (Winter and Perlman 2021:4), so broadening the scope would be a logical next step.

Other predicted correlations that could not be confirmed were those between distributive and full and echo reduplication. There seems to be some evidence for a correlation between distributive and full reduplication, but a correlation with echo reduplication may be precluded by the stronger association with similar. Thus, a more complex model that takes into account the relative strength of hypothesized iconicities may be helpful as well.

5.3 Implications

In line with the work of Lǐ and Ponsford (2018) on predicative reduplication, my analysis shows that form-meaning pairings can be iconic in different, complementary ways. Furthermore, I extended Lǐ and Ponsford’s methodology to include partial and echo reduplication, showing that each formal type has, to some extent, its own semantic realm. In this respect it is most difficult to distinguish partial reduplication, which can not only be freshly coined but also arise from a grammaticalization process.

Most modern authors assume some form of a complexity iconicity to explain the conglomeration of iconic and apparently counter-iconic functions in reduplication (“complex form reflects complex meaning”; see e.g. Abraham 2005; Stolz 2007; Stolz et al. 2011; Mattes 2014; Mattiola and Barotto 2023). This article has shown that at least some features can be linked to specific formal types of reduplication, which indicates that a more precise approach is needed. Thus, while I agree that so-called counter-iconic functions are only apparently counter-iconic, I do not trace back all meanings to a single highly abstract iconicity, but define several more fine-grained iconicities which can together account for the spectrum of attested meanings. Unlike an approach based on one highly abstract iconicity, this generates hypotheses that can be falsified or confirmed using a probability sample, namely by considering correlations between semantic features and formal properties.