A network analysis of the semantic evolution of ‘fruit’ and ‘stone’ in Tibeto-Burman languages

2.1 Dataset

A sample of 58 + 68 languages are collected on the basis of Sagart et al. (2019) to study the colexification patterns of ‘fruit’ and ‘stone’, respectively (i.e. see Appendix 1). Data collection and curation followed the following steps:

1. By querying the concepts ‘the fruit’ and ‘the stone’ in the database of Sagart et al. (2019),^[4] (https://dighl.github.io/sinotibetan/), two lists of languages containing the concepts ‘the fruit’ and ‘the stone’ in a sample of 117 Tibeto-Burman languages are compiled. The two lists are comprised of 19 cognate sets involving the lexeme ‘the fruit’ and 26 cognate sets involving the lexeme ‘the stone’ within the family of Tibeto-Burman. They are used as the starting points of the analysis. 104 languages are identified containing the concept ‘the fruit’ and 116 languages containing the concept ‘the stone’.

2. To ensure that word forms in a particular semantic network are derived from the same etymon, we use the database of STEDT^[5] (Matisoff 2015) (https://stedt.berkeley.edu/∼stedt-cgi/rootcanal.pl) as the reference point.^[6] Based on the word forms for the concepts ‘the fruit’ and ‘the stone’ in the two lists of languages, we searched words containing the morpheme(s) glossed as ‘fruit’ and ’stone’ in STEDT for each sampled language. Concepts that are colexified with ‘fruit’ and ‘stone’ in all sampled languages are gathered through this procedure. Since compounding is a productive strategy of word formation in the Tibeto-Burman family, the lexemes ‘fruit’ and ‘stone’ can be monosyllabic or disyllabic. The selection criteria is: if a word shares at least one syllable with the word ‘the fruit’/ ‘the stone’, it is included in the dataset for further analysis. By way of illustration, to compile the colexified word forms of ‘fruit’ in Tibetan (Batang), we first search the word(s) glossed as ‘fruit’ in STEDT. ‘Fruit’ in this language is a compound xhĩ ⁵⁵ thoʔ ⁵³ in which the first morpheme xhĩ ⁵⁵ is derived from the etymon ‘TREE/WOOD/FIREWOOK’(#2658) in PTB. Word forms in STEDT containing either xhĩ ⁵⁵ or thoʔ ⁵³ are then obtained. Table 8 shows the word forms containing the form xhĩ ⁵⁵ in Tibetan (Batang).

Table 8:

Word forms colexified with xhĩ ⁵³ in Tibetan (Batang).

1. Manually annotate the proto-form and the etyma tag of the colexified lexical forms of ‘fruit’ and ‘stone’ for each language on the basis of the etyma tag (e.g. #2658 in Table 8) and the corresponding proto-form in STEDT.

2. Among the languages that are initially collected from step 1, languages in which the cognate set of ‘fruit’ and ‘stone’ is undetermined (Cog ID = 0) are excluded, except for languages with apparent cognates (e.g. Ngwi-Burmish languages all have the cognate of *sey for ‘fruit’). Consequently, the sample contains 58 languages for ‘fruit’ and 68 languages for ‘stone’.

3. Since colexification network only concerns polysemy (i.e. senses that are related), it is necessary to filter out homonyms (i.e. senses that are unrelated). A practical solution is to filter the spurious links on a colexification network, as they appear in only one or two languages (Di Natale and Garcia 2023; List et al. 2018; Rzymski et al. 2020). For this reason, we first sort the concepts by frequency in the language sample and remove the concepts that only occur in one language in the entire dataset.

Table I and Table II in Appendix 1 present the sampled languages and their subgroups, the word forms of ‘fruit’ and ‘stone’ in each language, the cognate set ID and the etymon tag in STEDT of each word form, and the number of concepts that have the same form as ‘fruit’ and ‘stone’ (i.e. colexify) in a particular language.

2.2 A network-based methodology

Network-based approaches, increasingly popular due to recent advancements in computational science and graph theory, have found wide application in diverse fields such as physics, psychology, and knowledge engineering (Castro and Siew 2020; Kenett and Faust 2019; Siew 2020). In semantics, these methods have proven especially beneficial. A semantic network, as outlined by Steyvers and Tenenbaum (2005), consists of nodes representing words/concepts, and links portraying the relationships between them. Semantic networks constructed within or across languages have thus facilitated investigations into a range of phenomena including semantic changes across the lifespan and bilingualism (Borodkin et al. 2016; Wulff et al. 2022), organization of nouns and verbs in the mental lexicon (Qiu et al. 2021), and cross-linguistic colexification (List et al. 2013).

There are several benefits to using semantic networks to analyze word meaning. First, network representation provides a powerful tool for visualizing semantic relations and dynamics, allowing researchers to easily identify patterns and structures that may not be apparent in other forms. In addition, network analysis techniques can help identify important properties and behaviors of semantic networks mathematically using graph theory. For example, network analysis can reveal the properties of individual nodes, as well as clusters, sub-networks, and the overall network. One of the most important node-level properties is the degree, which refers to the number of links a node has. Words/concepts with a high degree are highly connected to other words, and thus have a greater influence on the overall structure of the semantic network. Important cluster/network-level properties include the clustering coefficient (CC; a measure of the local density of links) and average shortest path length (ASPL; the average number of steps along the shortest paths for all pairs of nodes). When semantic networks exhibit high CC and low ASPL, they are considered to have a “small-world” structure. This type of structure is characterized by highly connected clusters of nodes, or “communities”, that are themselves relatively well connected to one another. At the same time, the network as a whole maintains short path lengths between any two nodes, allowing for efficient communication and the spread of information across the network (for a review, see Siew et al. 2019).

3.1 Semantic classes colexified with ‘fruit’ and ‘stone’

Using the sampling procedure outlined in §2.1, 104 and 99 concepts are identified in the sampled languages that share at least one morpheme (i.e. colexify) with the lexeme ‘fruit’ and ‘stone’, respectively. Those concepts constitute the crosslinguistic colexification networks of the two nouns in Tibeto-Burman. In the initial sampling of languages, 58 languages are selected for ‘fruit’ and 68 for ‘stone’. Due to the existence of languages that do not colexify with any concept, 52 and 54 languages are eventually included in the network estimation of 'fruit' and 'stone', respectively.^[8]

Concepts colexified with ‘fruit’ are subdivided into 11 semantic classes: varieties of fruits, seeds/grains, body parts, small animals, small round objects, plants/plant parts, wood/tree, human/god, process/property, classifiers, and miscellaneous. They can be further grouped into 7 macro semantic classes.^[9] Concepts that are found in at least one language in the sample are included for analysis (i.e. Appendix 2, Table III). Table 9 presents the most frequent 10 concepts colexified with ‘fruit’ in the 58 sampled languages.

Concepts colexified with ‘stone’ are also divided into 11 semantic classes: varieties of stones, tools, fabrics, body parts, small animals, small round objects, direction, process/property, human/god, numeral classifiers, and miscellaneous.^[10] Table IV in Appendix 2 presents the concepts that are found in at least one language in the sample. Table 10 presents the most frequent 10 concepts containing the morpheme ‘stone’ in the 68 sampled languages.

3.2 Colexification networks of concepts

The network approach (§2.2) is applied to concepts colexified with ‘fruit’ (§3.2.1) and ‘stone’ (§3.2.2) to discover the structure of concepts in terms of their colexification similarity. The colexification networks of ‘fruit’ and ‘stone’ are characterized by Figure 1.

Figure 1:

Colexification Network of ‘Fruit’ (top panel) and ‘Stone’ (bottom panel). In this visualization, varying text colors denote different semantic classes (§3.1) of concepts. Node colors signify distinct clusters within the network, illustrating how concepts are grouped based on their colexification patterns. The size of each node corresponds to the degree of the concept, with larger nodes indicating concepts that have a higher degree of colexification with other concepts.

Table 11 presents the network measures for the colexification networks for ‘fruit’ and ‘stone’. Results of the one-sample z-tests between the CC and ASPL of the colexification networks and their corresponding random networks demonstrated that the two colexification networks were significantly different from the simulated random networks (p’s < 0.001), suggesting that the two networks were structurally meaningful and not simply the result of chance. In addition, the two networks exhibited small-world properties (i.e., CC ≫ CC_random and ASPL ≥ ASPL_random), which was further supported by the degree distribution, as shown in Figure 2, where a few nodes have a very high number of connections, while most nodes have only a few connections.

Figure 2:

Relative Frequencies of Node Degrees in the ‘Fruit’ (top panel) and ‘Stone’ (bottom panel) Networks. The bar charts show the distribution of node degrees in the two colexification networks, where the x-axis (i.e. degree) represents the number of connections a node has, while the y-axis indicates the relative frequency of each degree. The skewed distribution, with a majority of nodes having few connections and a few nodes having many, is typical of small-world networks.

3.3 Network stability

The results from Levene’s tests indicated significant differences in the variances for both the CC and ASPL between the ‘stone’ and ‘fruit’ networks. For CC, the test yielded an F-value of 420.48 (p < 0.001), suggesting a significant higher variability of clustering in the ‘stone’ network (SD = 0.03) compared to the ‘fruit’ network (SD = 0.01). Similarly, for ASPL, the test resulted in an F-value of 6.91 (p < 0.05), reinforcing the conclusion of higher variability in the ‘stone’ network (SD = 0.46) than in the ‘fruit’ network (SD = 0.40).

Further, Welch’s t-tests were performed to compare the mean differences of the CC and ASPL between the networks. The t-test for ASPL showed a mean of 4.18 for the ‘stone’ network and 4.66 for the ‘fruit’ network, with a t-value of -24.66, indicating a statistically significant difference in the ASPL between the two networks (p < 0.001). For the CC, the mean values were 0.46 (‘stone’) and 0.48 (‘fruit’), with a t-value of -19.77, also signifying a significant structural difference in terms of clustering (p < 0.001).

These findings suggest notable differences in the stability of network construction between the ‘stone’ and ‘fruit’ networks, with the latter being significantly more stable. The significant variations in both the CC and ASPL imply that the two networks are structurally dissimilar, with the ‘stone’ network exhibiting a shorter average path length and lower clustering coefficient compared to the ‘fruit’ network.

4.1 Semantic extension from ‘fruit’ to 3D-classifiers

The 58 sampled languages can be divided into two classes: languages have a 3D-classifier and languages do not develop a 3D-classifier. Table 14 presents the 9 languages in our sample that employ a 3D-classifier derived from ‘fruit’, most of them are from the Ngwi subgroup.

The language clusters^[13] in Figure 3 on the basis of the similarity of colexification networks across the sampled languages reveal that the colexification pattern of ‘fruit’ is strongly associated with the subgroup of a particular language (χ² = 131.85, df = NA,p < 0.001, Cramer’s V = 0.80). It clearly shows the clustering of two groups of Ngwi languages (i.e. Cluster 1 and 2), indicating Ngwi languages are identical in their colexification networks of ‘fruit’. 7 out of the 9 languages employing at least one 3D-classifier are from these two clusters of Ngwi.^[14] There is a split between Northern (Yi) and Southern (Hani) Ngwi, and Central Ngwi (i.e. Lahu_Lancang, Jinuo, Yi_Sani, Lisu) in between. We speculate there exists some hidden layers to account for the split that is not relevant to the derivation of 3D-classifiers.

Figure 3:

Language Networks of ‘Fruit’. In this figure, distinct text colors correspond to different subgroups for languages. Node colors designate various clusters, showcasing how languages group together based on shared colexification patterns for ‘fruit’. The size of each node represents the degree of the language, with larger nodes indicating languages that have a higher degree of colexification with other languages in the network.

The most salient pattern that colexifies ‘fruit’ and 3D-classifiers in the network analysis (§3.2.1), as characterized in (1), is the shared pattern of Ngwi languages in our sample. Due to the strong colexification of ‘fruit’, compounds for fruits, and small round body parts (i.e. Cluster 3 (varieties of fruits) and Cluster 1 (small round body parts) are the two central concept clusters in the network of ‘fruit’), semantic extension from ‘fruit’ to 3D-classifiers is most likely through the mediation of compounds for varieties of fruits and small body parts.

Diachronically, (1) indicates that the noun ‘fruit’ does not directly derive classifiers. For both languages with and without a classifier, ‘fruit’ usually first developed into a CT in compounds marking varieties of fruits. However, not all languages that involve the CT ‘fruit’ in the fruit compounds have developed a classifier. The CT ‘fruit’ tends to be present in compounds denoting small round body parts from Cluster 1 before developing the use as a classifier.

Table 15 presents four types of languages in our sample that exhibit variation in (1), resulting in the presence/absence of a 3D-classifier.^[15]

The type I languages follow the path in (1) and are primarily from the Ngwi subgroup. Four Ngwi languages that colexify ‘fruit’ with the Cluster 3 fruit compounds as well as the Cluster 1 body part compounds also possess a 3D-classifier. The only exception is Yi_Weishan, which satisfies both conditions in (1) but does not contain a colexified 3D-classifier.

The type II languages colexify ‘fruit’ with the fruit compounds in Cluster 3 but not the body parts from Cluster 1. As expected, they do not develop a 3D-classifier. Languages outside the Ngwi subgroup may colexify ‘fruit’ with fruit compounds but do not contain a classifier. This is probably due to that the ‘fruit’ morpheme is not colexified with body part terms, as demonstrated by Bola_Luxi and Burmese_Rangoon, or the colexified body part term is from clusters other than Cluster 1 (i.e. Cluster 4 or Cluster 6), as in as Dulong, Jingpho, Bantawa, and Chepang.

The type III languages colexify ‘fruit’, fruit compounds and 3D-classifiers but do not colexify ‘fruit’ with body parts from Cluster 1. Those languages involve Lisu, Hani_Mojiang, and Hani_Lüchun. However, taking a closer look at those Ngwi languages, they should not be excluded from the Type I languages. We found an array of Cluster 1 body part compounds that are colexified with ‘fruit’ in different sources.^[16] Therefore, the path in (1) still holds for the majority of Ngwi languages.

The type IV languages do not colexify ‘fruit’ with fruit compounds in Cluster 3 nor body parts from Cluster 1 but have developed a 3D-classifier. Naxi and Tani_Bokar in our sample exhibit this pattern.

There are of course other minor paths that derive a 3D-classifier in addition to (1). In Type IV languages, classifiers in Cluster 4 are not derived from ‘fruit’-related concepts but seem to be directly derived from body parts. For example, Naxi has the classifier ly ³³ classifying objects like grains, eggs, and bowls. However, ly ³³ does not colexify with any compound meaning ‘fruit’ but rather with body parts like ‘finger’ and ‘kidney’. Tani_Bokar uses the classifier pɯ rɯ to classify bowls, which is originally derived from *pɯ ‘egg’ in Proto-Tani (Sun 1993). Like Naxi, it is not found colexified with varieties of fruits but rather is found in ‘gall’.

Based on the above discussion, the presence of a numeral classifier in the colexification network of ‘fruit’ is largely due to that the language is from the Ngwi subgroup. Both Type I and Type III Ngwi languages derive a 3D-classifier via the path in (1). The concept ‘fruit’ may derive the meaning of classifier as a result of shared innovation of Ngwi languages. This is congruent with Bradley (1979), in which the proto *si ² is reconstructed for both ‘fruit’ and the classifier ‘CLF: fruit’ in Proto-Ngwi.

4.2 Semantic extension from ‘stone’ to 3D-classifiers

The 68 sampled languages can be divided into two groups depending on the presence/absence of a 3D-classifier. Table 16 presents the 7 languages in our sample that employ a 3D-classifier derived from ‘stone’. In contrast to ‘fruit’, which is highly biased to Ngwi languages, languages deriving a 3D classifier from ‘stone’ are distributed over 6 subgroups. The subgroup (χ² = 108.36, df = NA, p = 0.04) of a particular language is marginally associated with the colexification pattern of ‘stone’.

According to Figure 1, no salient path from ‘stone’ to 3D-classifiers can be identified in the concept network of ‘stone’, since the sub-networks of classifiers are diverse and somewhat isolated from the stone-related compounds. Compounding involving ‘stone’ as an intermediate stage in the derivation of classifier is not as productive as (1) in TB languages.

Figure 4 shows that at least two groups of languages (i.e. Cluster 1: Garo, Dulong, Karen, Bai_Jianchuan, Nung, Daofu; Cluster 4: Tibetan_Written) that exhibit distinct colexification patterns of ‘stone’ have independently developed 3D-classifiers. Further splits can be made within Cluster 1 once scrutinizing this cluster. It is due to the high variability of the network of ‘stone’, indicating that the path deriving a 3D-classifier from ‘stone’ is very unstable (§3.3). It follows that TB languages exhibit dissimilar colexification networks of ‘stone’, which are ineffective in predicting the presence of classifiers in a particular language. Below we will take a closer look at the 7 languages that possess a 3D-classifier and demonstrate the inability to associate the colexification pattern of ‘stone’ with a 3D-classifier.

Figure 4:

Language Networks of ‘Stone’

Table 17 displays the 7 languages that colexify ‘stone’ and 3D-classifiers. ‘Stone’/ ‘stone’-related compounds and 3D-classifier exhibit relatively direct links and shorter path in derivation as the colexified ‘stone’ compounds are confined to a small number of concepts (most prominently ‘flint’). Nung, Dulong, Jianchuan_Bai, Karen, and Written Tibetan all colexify ‘stone’ with compounds for varieties of stones, while the root ‘stone’ is not found in ‘stone’-related compounds in Garo and Daofu. The path ‘Noun – CT in compound nouns – 3D-classifier’ holds for most of those languages. Nevertheless, the concept network result in §3.2.2 shows that ‘stone’ concepts of the same semantic class are not strongly colexified. Notably concepts pertaining to the semantic class of ‘small round objects’ in Table 17 are quite diverse across languages and hence no intermediate stage concerning compounds for ‘small round object’ should be posited between ‘stone’-related compounds and 3D-classifiers. It implies that the overall colexification pattern of ‘stone’ is not strongly related to the presence of a 3D-classifier, as the classifier is more directly derived from ‘stone’ and a limited number of ‘stone’-related compounds.

Combining the network instability (§3.3), the language clustering in Figure 4, and the language-specific data in Table 17, 3D-classifiers seem to be derived from ‘stone’ independently in individual languages and cannot be uniformly accounted for by semantics.

5.1 Generic-specific compound nouns in the semantic extension towards 3D-classifiers

There is a strong tendency of grammaticalization of numeral classifiers from nouns in Tibeto-Burman and worldwide (Aikhenvald 2000; Corbett 1991; Evans 2022; Huang 2022; Post 2022; Grinevald 2000, 2002; Seifart 2010). Results from the colexification networks of concepts in this study (§3.2) point to that compound nouns serve as a critical semantic class in this process. Among the TB languages that have developed 3D-classifiers classifying small round objects such as ‘egg’, ‘grain of rice’ and ‘stones, rocks’, 3D-classifiers are indirectly colexified with the nouns for ‘fruit’ and ‘stone’ because 3D-classifiers and ‘fruit’/ ‘stone’ are distributed in distinct clusters in both networks. The centrality and high degrees of compound nouns containing the morpheme ‘fruit’/ ‘stone’ in both concept networks strongly indicate that compounds are colexified with more concepts than 3D-classifiers. It follows that compounding is an intermediate stage in the derivation of 3D-classifiers from nouns in Tibeto-Burman langauges. This finding can be empirically supported by crosslinguistic data.

Classifier systems typically emerge in two distinct contexts (Little et al. 2022): the context of counting individual items which are of particular cultural importance (Bisang 1999:158), such as Chinese and Japanese; the context of taxonomic or meronomic compounding process, which is prominent in Tai, Hmong-Mien, and Tibeto-Burman languages (Bisang 1999; DeLancey 1986; Enfield 2004; Vittrant and Allassonnière-Tang 2021). ‘Class noun’ (or class term) as a part of the noun root is conceived as an important stage in the process of deriving a “classifier-for-noun” (Little et al. 2022). (2) through (4) illustrate distinct types of classified nouns in Mandarin Chinese, Hmong, and nDrapa.

The general classifier gè in Classical Chinese was not evolved from any generic terms in compounds but initially used as a general classifier, whose occurrences as a numeral classifier can be dated back to as early as Han Dynasty (Zhang 2012). In Hmong, by contrast, there are class terms like ntoo ‘tree’, txiv ‘fruit’, tub ‘son’, etc. (Bisang 1999: 167), which serve as generic nouns specifying a superordinate level concept. They are the lexical origin of numeral classifiers, as in (3). The small set of numeral classifiers in the Tibeto-Burman language nDrapa are also grammaticalized from compound nouns that hold a generic-specific relationship between the classifying morpheme and the host noun. The ji coda of the noun ‘apprentice’ in (4) means ‘man/human’ and serves as a class term specifying the superordinate class of láɕheʂtsʊ ‘pupil’. It further developed into a general classifier jî (Huang 2022: 231).

Findings of the present quantitative study support the claim that compound nouns play a critical role in the grammaticalization of TB classifiers. We propose the path in (5) to characterize the semantic extension of 3D-classifiers from the nouns for ‘fruit’ and ‘stone’, in which the lexeme ‘fruit’ or ‘stone’ serve as a class term (CT) in compound nouns before turning into a pure classifier. The CT and the other part of the compound hold a generic-specific semantic relationship.

This grammaticalization cline is not only widespread in TB languages that have rich classifiers, such as Ngwi-Burmish and Tani languages (Bradley 2012; Post 2022), but also common in languages with limited number of classifiers, as in nDrapa (Qiangic, see Shirai 2022; Huang 2022).

‘Fruit’ first developed into a bound morpheme (CT) in compounds for varieties of fruits, such as ‘pear’, ‘peach’ and ‘walnut’. The CT ‘fruit’ was extended to compounds of the same shape as fruits and finally became a shaped-based classifier. As illustrated by Figure 1, ‘CLF: eggs’, ‘CLF: stones, rocks’, ‘CLF: bowls’ and ‘CLF: grain (of rice)’ all have direct links with compound nouns for types of fruits or small round objects, but none of them are directly colexified with the noun ‘fruit’. Taking Lisu (Central Ngwi) as an example. Table 18 presents words colexified with the lexeme ‘fruit’, including the classifier ‘CLF: grain (of rice)’ (Sun 1991;Huang 1992). sɯ ³¹ initially appeared in the disyllabic generic noun for ‘fruit’ (Stage 1); it then developed as a class term in compound nouns for types of fruits (Stage 2); in Stage 3, sɯ ³¹ occurred in compounds for grains, seeds, body parts, and other round-shaped objects; it finally became a sortal classifier for grains (Stage 4).

In the case of ‘stone’, as shown in Figure 1, it is not directly linked with the classifiers ‘CLF: stones, rocks’, ‘CLF: grain (of rice)’, and ‘CLF: eggs’ but rather derived those classifiers via compound nouns for varieties of stones and other round things. The compounds for ‘flint (to make fire)’, ‘flight (of steps)’, and ‘coal’ are most important nouns that derived classifiers in an array of languages (§4.2.2). The morpheme ‘stone’ appears as a class term in those compounds before turning into a classifier. Table 19 presents the colexified concepts of ‘stone’ in Dulong (Nungic, Huang 1992). The noun root luŋ ⁵⁵ ‘stone’ (Stage 1) first appears in the compounds for different types of stones and rocks (Stage 2), holding a specific-generic relationship between the basic and the superordinate level concept; luŋ ⁵⁵ then developed into a class term accompanying inanimate round objects that are small in size (Stage 3). Finally, it became a 3D-classifier (Stage 4).

Shape is considered the most important semantic parameter in noun categorization, including numeral classifiers (Adams and Conklin 1973; Aikhenvald 2000; Senft 1996). The semantic colexification pattern of ‘fruit’ and ‘stone’ of the sampled Tibeto-Burman languages demonstrates a crosslinguistic tendency that before this parameter comes into play, the “shape” lexeme is a taxonomical classifying morpheme that classifies nouns according to their superordinate level category.

5.2 The effectiveness of colexification patterns in predicting 3D-classifiers

‘Fruit’ and ‘stone’ diverge in the mode of deriving 3D-classifiers from compounds. The significant difference in their network structure and stability (§3) and the language-specific data (§4) strongly suggest that only ‘fruit’ but not ‘stone’ exhibits a stable and salient path in semantic extension from nouns to 3D-classifiers. Only the colexification pattern of ‘fruit’ but not ‘stone’ can effectively predict the presence/absence of a 3D-classifier.

The most salient path of semantic extension from ‘fruit’ to 3D-classifiers in Tibeto-Burman experiences the 4 stages characterized in (1) in §4.1 (i.e. repeated below), as illustrated by the Lisu example in Table 18.

(1) is a path attested in most Ngwi languages in our sample. Ngwi languages that developed 3D-classifiers from ‘fruit’ tend to colexify ‘fruit’ with the body part compounds ‘testicle’, ‘fingernail’, ‘ankle’, ‘claw’, ‘bladder’, ‘liver’, and ‘throat’ (Cluster 1) in Stage 3. A 3D-classifier is commonly absent in languages in which ‘fruit’ does not colexify with those body part concepts (i.e. Type II languages in Table 15). An implicational universal can be formulated to predict the presence/absence of a 3D-classifier based on their distributions in TB (i.e. Table 15):

(6) If a TB language colexifies ‘fruit’ with the ‘fruit’ CT in fruit compounds from Cluster 3 and the shape morpheme in body part concepts from Cluster 1, the language tends to have a 3D-classifier.^[17]

The four types of languages investigated in Table 15 (§4.1) empirically support the prediction in (6). The path in (1) turns out to be fairly in Tibeto-Burman that accounts for half of the languages that possess a 3D-classifier and is highly skewed toward Ngwi.

‘Stone’, to the contrary, does not exhibit any crosslinguistic salient path of semantic extension like ‘fruit’. The high variability of the network of ‘stone’(§3.3) and the weak colexification patterns of the ‘stone’-related concepts (§3.2.2), together with the language clustering in Figure 4 and the language-specific data in Table 17 (§4.2), all indicate that the path deriving a 3D-classifier from ‘stone’ is very unstable. The shorter average path length and the direct links between individual ‘stone’ concepts (e.g. ‘flint’) and 3D-classifiers in the network of ‘stone’ (§3.3) suggest that the semantic extension ‘stone’ – ‘stone’related compounds- 3D-classifiers is incidental. Such extension is not semantically motivated as ‘fruit’ but rather the result of semantic extension of invidual nouns.

Indeed, the evolutionary situations from ‘stone’ to classifier are incidental and language-specific. Among the 7 languages that develop a 3D-classfier from ‘stone’, only Garo (Bodo-Garo) and Karen (Karenic) have reconstructed classifiers in the proto languages. In Karen (Karenic), lø ³¹ ‘stone’ is colexified with the classifier phlø ³¹ , which classifies stones, eggs, and grain of rice (Huang 1992). This classifier has already existed in the Proto-Karen (i.e. *phloŋ ^B ‘CLF: stones/rocks’), which was colexified with the proto form for stones (i.e. *loŋ ^B ‘stone/rock’) (Luangthongkum 2013). In other languages, the colexification of ‘stone’ and classifier is absent in the their corresponding proto languages. Classifiers are a part of Proto-Bodo-Garo (Wood 2008), Proto-Tani (Post and Sun 2017), and Proto-Ngwi (Bradley 1979). But none of the classifier roots in those proto languages are related to ‘stone’. Some languages have experienced replacement of the 3D -classifier in history, resulting in the colexification of ‘stone’ and 3D-classifier in modern languages. Proto-Boro-Garo (PBG) has a handful of numeral classifiers (Wood 2008), among which the PBG classifier of round objects has the same proto form as the lexeme of ‘fruit’ (i.e. *thái)^[18] but later in some languages this classifier was replaced by the lexeme of ‘stone’ (*roŋ-). This is what has been observed in Garo (Wood 2008:75).

Given the above facts, we may conclude that though ‘fruit’ and ‘stone’ both derive classifiers via an intermediate stage of compound nouns, their colexification networks strongly imply two distinct modes of deriving classifiers from nouns. The colexification network of ‘fruit’ but not that of ‘stone’ can effectively predict the occurrence of a 3D-classifier in a particular language, which is highly correlated with the subgroup of the language.

5.3 Body part concepts in noun categorization

The essesntial status of body part compounds in the derivation of 3D-CLF in TB highlights the importance of body part concepts in noun categorization. The close relationship between body parts and the small round shape has been characterized in many noun categorization systems. For example, the gender assignment in New Guinea (Aikhenvald 2012) is associated with the shape of the body part (i.e. round or long). In Nilo-Saharan languages, body parts typically categorize nouns in terms of their shape (Blench 2015). Body parts appear to be a common lexical source for shape-based numeral classifier in Asian classifier languages (Aikenvald 2000; Bisang 1996) as well as noun classes/gender in other parts of the world, as in Bahnar (Adams 1989), Kana (Ikoro 1996), and Gumuz (Blench 2015) in Africa, and Palikur in Amazon (Aikhenvald and Green 1998). Findings of this paper suggest that body part compounds serve as a critical stage in the development of 3D-classifiers in Tibeto-Burman languages. However, it is noteworthy that only a specific subset of body part concepts, including ‘testicle’, ‘fingernail’, ‘ankle’, ‘claw’, ‘bladder’, ‘liver’, and ‘throat’(Cluster 1), are relevant to the derivation of a 3D-classifier. There is another small set of body part concepts, such as ‘eye’, ‘heart’, and ‘naval’, in Cluster 6 that do not play a role in the derivation of 3D-classifiers. Colexification with those body parts is not a sufficient condition of the presence of a classifier.