2.1 General ideas

When assembling comparative word lists for the purpose of phylogenetic reconstruction, the major problem imposed by language families in which partial cognacy is frequent is that it often becomes very difficult to find clear-cut criteria to assign words to cognate sets. In abstract terms, if one language expresses a concept ‘X’ with a compound word a-b and another language expresses the same concept with a compound word a-c, there are two possibilities: one could either argue that the two words are to be judged as cognate, given that they have one cognate morpheme a in common; or one could argue that they are not cognate, given that they differ due to their respective morphemes b and c, which are not cognate. The complexity increases when more words are brought to the comparison and can easily lead to cases where the decision to assign all words which share at least one common morpheme to the same cognate set yields situations in which our hypothetical word a-b would be cognate with a-c and a-c would be cognate with d-c, but d-c would no longer share any common element with a-b.

The two most straightforward approaches to assigning words to cognate sets when their partial cognate sets are known have been called “strict” and “loose” cognate coding in previous work (List, 2016; List et al., 2016). In the strict case, only those words which are cognate with respect to all of their morphemes are assigned to the same cognate set. An example for this coding is the study on Chinese dialect evolution by Hamed and Wang (2006). In the loose case, a network of all words is constructed in which words correspond to nodes and links between nodes are drawn whenever two words share at least one cognate morpheme. After the network has been constructed, all words that belong to a connected component in the network are assigned to the same cognate set (Hill and List, 2017). An example for this coding procedure can be found in the study by Satterthwaite-Phillips (2011). Each approach has its advantages and disadvantages. While strict coding may easily increase differences between language varieties, giving the incorrect impression that there is a huge amount of linguistic variation in a given language family, the loose coding practice is unsatisfying as it may easily result in cognate sets consisting of word pairs that do not have a single cognate morpheme in common.

Assuming that partial cognates have been identified, an additional way to code the data in phylogenetic analyses would consist in ignoring the word level and coding the partial cognate sets directly. This technique, however, would contradict the important criterion of character independence, since individual morpheme cognate sets have not been evolving alone, but together with the words in which they appear. Since character independence is one of the basic criteria upon which phylogenetic models are built, introducing character dependencies may not only impact phylogenetic reconstruction (Felsenstein, 1988: 446), it will also make the results extremely difficult to interpret, since we ultimately want to understand how whole words evolve during language evolution, not how certain morphemes are gained and lost.

In order to avoid counting words which do not share a single cognate morpheme as cognate, Sagart et al. (2019) annotate their cognate sets in such a way that all words assigned to the same cognate set must at least have one morpheme in common. While this coding practice is beyond doubt more principled than the strict or the loose coding practices mentioned before, it has the disadvantage that it cannot be automatically checked. Sagart et al. (2019) make use of alignment analyses in order to make sure that there is a common morpheme in large cognate sets, but since they do not mark partial cognates in their data, it is not trivial to check all of their codings automatically. As a result, it is possible to check the consistency of their cognate annotation, but it is not easy to do so, since one has to go manually through each entry.

It is never trivial to decide whether overall cognacy for a set of words should rely on the presence of one single morpheme shared by all words or the presence of several words. As an example, consider the concept ‘sun,’ which many Austronesian languages lexify as ‘eye of the day,’ with the form for ‘day’ often being equivalent to the original word for ‘sun’ (Starostin, 2013: 121–123). Should we say that in a language which retains the original word for ‘sun’ this is cognate with a word in a language which shows the motivation ‘eye of the sun/day,’ or should we rather say that the latter is an innovation and reflects a clear case of lexical replacement? We think that this question cannot be clearly answered, but depends on the language family in question and our knowledge about it. The problem can therefore not be resolved by a computational approach alone.

While it is not possible to design a straightforward algorithm that would make the cognacy decisions in our place, it is, however, possible to insist on a more explicit annotation of lexical cognacy data that would reflect the individual decisions on cognacy taken by individual scholars. The solution we propose for this task is to make use of morpheme glosses, as shown above for the Tupían data in Table 2. Morpheme glosses were first proposed by Hill and List (2017) and further developed by Schweikhard and List (2020). We extend this work by adding a new aspect to the analysis, insofar as we mark the morpheme or the morphemes which we consider as salient with respect to the history of the word in question. Under saliency we understand the potential of one or more morphemes to reflect the major evolutionary processes of the words in which they occur.

As an example, consider the words for ‘head’ in Tupían languages, which can be roughly divided into those words that denote head directly, such as Cinta-Larga [antar], words that involve a morpheme for ‘hair,’ such as Awetí [ʔap-ut], and words that contain a morpheme that means ‘round,’ such as Akuntsu [a-nãm] (with [a] glossed as ‘round’). One potential analysis of these partial cognates would be to take ‘round’ as the salient morpheme and to assume that it reflects an innovation in the language family, which was later diversified, leading to various subtypes that can or should be ignored in a phylogenetic analysis. Another possibility would be to say that the specific combination of ‘round’ and ‘bone’ should be treated as the major innovation. In this case, Amanaye [a-kɨ] and Amondawa [a-kaŋ] would reflect one common innovation and therefore be treated as one cognate set, while the other words that contain a reflex of ‘round’ but no reflex of ‘bone’ would be kept apart. Table 3 illustrates the consequences of these two decisions regarding the saliency of the morphemes with respect to the evolutionary history of their words.

This idea of marking those morphemes in the morpheme glosses which one identifies as representative for the word history can be seen as a less restricted variant of the aforementioned strict conversion of partial cognates into cognate judgments on whole words. While the strict conversion takes all morphemes in a given word as equally important, our proposal to annotate which morphemes are salient and which are not allows scholars to exclude specific morpheme cognates from the equation. As a result, scholars can, for example, argue that a certain suffix occurs so frequently in a given data set that it does not play a significant role in deciding whether a word that has the suffix should be considered cognate with a word that lacks the suffix.

Morpheme glosses are a free annotation form that serves to describe the semantic motivation structure of a given word. The term “motivation” is based on Koch (2001) and is used by Hill and List (2017) and Schweikhard and List (2020) to denote the semantics underlying word formation processes. As an example, consider Mandarin Chinese shù-pí 树皮 ‘bark (of tree)’, which consists of the two morphemes shù 树 ‘tree’ and pí 皮 ‘skin.’ The semantic motivation underlying the compound is thus the metaphorical use of ‘skin’ to denote the cover of trees. Hill and List (2017) indicate these motivation structures in their tabular word list data with the help of an extra column in which individual morphemes of multimorphemic words are glossed.

As an example for this annotation practice, consider the example of words denoting ‘hatchet’ in six Mienic varieties (original data taken from Máo, 2004) given in Table 4. In this table, we can observe three distinct morphemes from which all six words are built. All words share one morpheme that means ‘knife’ in isolation (colored in red in the table), but in Daping and Dongshan, the reflexes dziu²² and ȡu⁴² appear at the end of the words, while they appear at the beginning in the other four varieties. The first morphemes in Daping and Dongshan are reflexes of Proto-Hmong-Mien *dzaŋ^A ‘firewood’ in the reconstruction of Ratliff (2010: 254), and the semantic motivation of the words in the two varieties is ‘firewood-knife,’ indicating that a hatchet is a specific kind of knife predominantly used for the preparation of firewood. In the remaining four varieties, where the morpheme for ‘knife’ appears at the beginning of the word, the second morpheme can be translated as ‘bent, crooked’ in isolation. Since most Mienic languages place the modifier after the modified, the semantic motivation for ‘hatchet’ is ‘bent knife,’ that is, a knife that has a bent form.

Once morpheme glosses have been added to a data set, the annotation of salient morphemes, that is, morphemes one deems representative for the whole history of the words, can be done in a very straightforward way by simply indicating the saliency along with the morpheme glosses. In our concrete annotation, this means that we add an underscore _ in front of each morpheme gloss which we consider as not salient. When later converting partial cognates to “full” cognates, we only extract those cognate sets whose morpheme glosses have been annotated as salient and then use the strict conversion procedure on these selected cognate sets.

As an example for this procedure, consider the words for ‘belly’ in five Hmongic languages in Table 5 (Chén, 2012: 599). All words show the same basic structure of being composed of a prefix with synchronically untransparent semantics and a main morpheme with the core meaning ‘belly.’ As can be seen from our partial cognate annotation (provided in the column “Partial”), we identify three distinct prefixes and two distinct morphemes for ‘belly,’ one going back to Proto-Hmong-Mien *chu̯ei^A in the reconstruction of Ratliff (2010), the other of an origin unknown to us. When computing strict cognate sets from the partial cognates, all words will be placed into distinct cognate sets, since none of the words coincide in all their morphemes. When using the procedure of loose cognate annotation, all words would be placed into the same cognate set, since they all form one big connected component, in which words containing a reflex of Proto-Hmong-Mien *chu̯ei^A, labeled belly/A in our morpheme glosses, are connected to the words with the reflex labeled belly/B via the prefix prefix/A, shared between Western Baheng and Chuanqiandian. Our procedure of salient cognate coding, on the other hand, deliberately ignores the prefixes—given that their presence or absence provides little evidence for the historical development of the words on which they occur, but rather points to largely language-specific processes of productive prefixation that are not well understood—and thus divides the five words neatly into two cognate sets, depending on the basic morpheme used to express the meaning of ‘belly.’

2.2 Specific ideas

The schema as presented in the previous section relies entirely on human judgment, and it is difficult—at least for the time being—to think of an automated approach to approximate human judgments. The reason is not the impossibility of finding alternatives to the strict and the loose practice of converting partial to full word cognate sets. As we will show in the following sections, we can easily implement a method that accounts for the cognate coding practiced by Sagart et al. (2019). The problem is that it is often not clear what should count as the best solution and that there is no real way to tell based on the data alone. In the following, we will nevertheless try to provide some general criteria that may help scholars in arriving at decisions in particularly difficult situations.

There are three major caveats when deciding about full word cognacy in multilingual word lists. First, when annotating cognates, scholars should try to avoid coding as cognates those cases that are highly likely to have evolved as a result of parallel independent evolution (i.e., avoid homoplasy). Second, one should try to make sure that the characters, that is, the cognate sets, are maximally independent (i.e., minimize character dependency). Third, one should make sure to identify cases of free or pragmatically conditioned synchronic variation and control for them systematically (i.e., control variation).

As an example for the first problem, that of parallel independent evolution or homoplasy, consider cases of lexical motivation in compounding (Koch, 2001). Words for ‘tears’ in Hmong-Mien languages are a good example, since as in many Southeast Asian languages, ‘tears’ tends to be expressed through a compound, of which one part in isolation is related to a word that means or originally meant ‘water’ (consider Mandarin Chinese lèi-shuǐ 泪水 ‘tears,’ which can be glossed as ‘tears-water’). In the Hmong-Mien languages, the other part of the compound is typically the same as the word for ‘eye,’ and the lexical motivation of ‘tears’ can thus be described as the ‘water’ of the ‘eye’ (Chén, 2012: 609). Unlike most Chinese dialect varieties, which tend to place the modifier before the modified in compounds, Hmong-Mien languages typically use the opposite order (‘water-eye’ instead of ‘eye-water’). In Sinitic, there are some exceptions of this rule in the south, which scholars tend to attribute to influence from the Hmong-Mien languages (Vittrant and Watkins, 2019), but we can find the opposite influence in some Hmong-Mien varieties as well. As a result, some Hmong-Mien languages lexify ‘tears’ as ‘eye-water,’ such as Zao Min mai⁵³-m²⁴ (mai⁵³ means ‘eye’ in isolation, going back to Proto-Hmong-Mien *mu̯ɛjH; and m²⁴ means ‘water,’ going back to Proto-Hmong-Mien *ʔu̯əm; see Chén, 2012; Ratliff, 2010), while the majority have a compound ‘water-eye,’ such as Western Qiandong ʔeu⁴⁴ me²² (ʔeu⁴⁴ is ‘water’ and me²³ is ‘eye’; Chén, 2012). Note that the morphemes in the words in Zao Min and Western Qiandong both go back to the same proto-forms, even if it is quite likely that the word for ‘eye’ was borrowed from Chinese. While it is trivial (despite the complex sound correspondences) to identify the morphemes in both words as cognate, it is far from trivial to decide on the cognacy of both words. One could assume that Proto-Hmong-Mien once had a compound ‘water-eye’ and that this compound was inherited by both Zao Min and Western Qiandong, and that the lexical motivation of the compound did not lose its transparency until Zao Min began to reverse the order of compound constituents from modified-modifier to modifier-modified, possibly under the influence of Chinese dialect varieties. The reversed word for ‘tears’ thus reflects some global innovation in the language which affected a large part of its lexicon. Another possibility, however, is to assume that the motivation underlying words for ‘tears’ in the Hmong-Mien languages is so obvious and general that we can easily assume that it could recur independently throughout the history of many languages. As a result, it would be wrong to say that the words as such are cognate, since one would assume that they were coined independently and therefore do not reflect shared innovations in the language family. With the knowledge we have at our disposal, we consider this case as undecidable. As a result, it seems best to ignore items like ‘tears’ when applying phylogenetic reconstruction methods to the Hmong-Mien language family in order to make sure that the phylogenetic signal is not contaminated by instances of parallel evolution.

As an example for the problem of character dependence, consider the analytical derivation of plural forms for personal pronouns in many Southeast Asian languages. While plural forms for personal pronouns tend to have an independent (suppletive) form in most Indo-European languages (compare German ich ‘I’ vs. wir ‘we,’ du ‘thou’ vs. ihr ‘you [pl.]’), many Southeast Asian languages derive plural forms from the singular forms by means of suffixation (Mandarin wǒ 我 ‘I’ vs. wǒ-men 我们 ‘we,’ nǐ 你 ‘thou’ vs. nǐ-men 你们 ‘you [pl.]’). As a result, the plural form can be regularly predicted from the singular form for most languages in which the plural is built analytically. However, many questionnaires for phylogenetic reconstruction in linguistics contain concepts for singular and plural personal pronouns, and so in these languages the corresponding characters for ‘I,’ ‘thou,’ ‘we,’ and ‘you (pl.)’ can no longer be considered to have evolved independently, since singular pronouns are reused to form the plural pronouns and all plural pronouns tend to share the same affix that derives the plural meaning.

When encountering these processes across all languages in a given data set, the only consequent way to deal with the cognate assignments is to code each morpheme only once, which would mean that one needs to modify the underlying questionnaire in such a way that only singular forms are used as the base forms, while plural forms of personal pronouns are collapsed into one single ‘plural’ category. If, however, not all plural forms are constructed analytically—as is the case for the Hmong-Mien languages, where some varieties have a regular plural suffix, similar to Mandarin Chinese (e.g., Jiongnai, a Hmongic language, has wa³¹ ‘I’ vs. wa³¹ kluŋ⁵³ ‘we’; Iu Mien, a Mienic language, has ʑe³³ ‘I’ vs. ʑe³³ wo³³ ‘we’), but some also have suppletive forms (Eastern Xiangxi, Hmongic, m³¹ ‘thou’ vs. ma⁵³ ‘you [pl.]’)—we recommend excluding plural forms directly from the analysis, since the independency of the characters cannot be guaranteed.

As an example for the problem of controlling variation, consider the phenomenon of affixation in the Hmong-Mien language family. In many Hmong-Mien languages, one finds a certain number of productive prefixes or suffixes which are typically used to derive nouns from a base form. Some of these derivations are mandatory, while some can be omitted, depending on the context. Thus, the word for ‘star’ in Xia’ao (Western Xiangxi, Hmongic branch of Hmong-Mien) will typically be elicited as qa⁰²-sin⁴⁴ (Chén, 2012: 145, 282), consisting of the prefix qa⁰²-, which derives inanimate nouns, and the noun sin⁴⁴, an early borrowing from Chinese xīng 星, which was pronounced as seŋ in the sixth century AD (Baxter, 1992). The use of the prefix, however, is not obligatory: it can be omitted, depending on the context (Chén, 2012: 145). When deriving cognate judgments for cases of this sort where free variation can be observed, we recommend first checking to ensure that the variation can be observed in all or most of the languages in a given sample, and if this is the case, excluding the longer forms from the data.

As we have tried to illustrate throughout this section: it is by no means trivial to deal with these questions, and we expect that the impact on phylogenies when adopting arbitrary solutions for cognate coding could be rather substantial. In order to address the problems in a straightforward manner, we suggest that scholars working with languages in which partial cognacy is a frequently recurring problem, resulting from abundant compounding and rich derivational processes, carry out a very close analysis of language-internal cognacy. Using morpheme glosses, it is possible to rigorously mark prefixes, suffixes, and the lexical motivation structures underlying compounds. Once this analysis has been carried out and partial cognates have been identified across languages as well as language-internally, thus taking both words with the same meaning and words with different meanings into account, scholars can carefully check individual semantic slots and try to identify whether any of the three problems discussed in this section applies. If this turns out to be the case, one should: (a) ignore the concepts that are expressed by words that are suspicious of parallel evolution due to frequently recurring patterns of lexical motivation (avoid homoplasy); (b) try to identify the phylogenetically important alternations when dealing with problems of character dependency and re-code the data accordingly (minimize character dependency); and (c) carefully study how words vary when being used in different contexts in order to handle problems resulting from language-internal variation (control variation).

3.1 Materials

The data set was originally published by Liú et al. (2007) and later digitized for this study by manually entering the data into text files. The data consists of 201 concepts translated into 19 Chinese dialect varieties (see Table 6) which provide at least one variety as a representative for each of the seven major subgroups proposed by Norman (1988: 181)—Mandarin (Guānhuà) 官话, Wú 吴语, Xiāng 湘语, Mǐn 闽语, Yuè 粤语, Gàn 赣语, and Hakka (Kèjiā) 客家—as well as one variety for each of the three subgroups which are often additionally proposed—Jìn 晋语, Pínghuà 平话, and Huī 徽语 (Yan, 2006). In order to guarantee the comparability of our data set with other data sets, we linked the concept list to the Concepticon reference catalog (https://concepticon.clld.org; List, Tjuka et al., 2022) and the language varieties to Glottolog (https://glottolog.org; Hammarström et al., 2021); see the supplementary material.

In the raw data, the translations for each concept in each variety are given in phonetic transcription and in Chinese characters (Liú et al., 2007). The latter are frequently used by Chinese dialectologists in order to mark etymologically related morphemes across different dialects (běn-zì 本字, literally ‘original characters’; see Mei, 1995). Although the Chinese character information on cognacy needs to be treated with some care, it is a good starting point for the annotation of cognate sets both across dialects and inside one and the same dialect.

Phonetic transcriptions in the original data set were standardized by converting the original transcriptions—which follow specific peculiarities as they are typically found in Sinitic varieties descriptions—to the transcriptions proposed by the Cross-Linguistic Transcription Systems (CLTS, https://clts.clld.org; List et al., 2021; see Anderson et al., 2018, for details on the CLTS system). This reference catalog is one of the core components of the Cross-Linguistic Data Formats (CLDF, https://cldf.clld.org; Forkel et al., 2018). The CLTS system can be seen as a narrower version of the International Phonetic Alphabet insofar as it resolves several of its ambiguities. For the conversion and segmentation of the transcriptions, orthography profiles (Moran and Cysouw, 2018) were used and all individual transcriptions were later manually checked.

Partial cognate sets were first automatically added to the data by employing the Chinese character readings, and later systematically refined using the interactive web-based EDICTOR tool for the creation of etymological data sets (https://digling.org/edictor; List, 2017, 2021). Morpheme glosses, following Hill and List (2017) and Schweikhard and List (2020), were manually added for all morphemes, based on the previously inferred partial cognate sets. In order to facilitate the reuse of the data, we used the CLDFBench software package (Forkel and List, 2020) with the Lexibank plugin (List, Greenhill et al., 2022) to convert the data to the tabular standards proposed by the CLDF initiative. The entire data set contains a total of 4,302 words, with 65.6 % of these being monosyllabic words and 34.4 % polysyllabic words.

The original data set of Liú et al. (2007) often contains multiple translations for the same concept in the same variety, and this can easily influence the results of phylogenetic reconstruction approaches. We therefore carefully excluded some of the translations which reflect specific colloquial registers. Following standard practice in phylogenetic reconstruction in historical linguistics, we also made sure to mark known borrowings in the data, relying on our own knowledge of Chinese dialect history as well as cases of borrowings annotated in similar data sets (Sagart et al., 2019). All decisions about which items were excluded or marked as borrowings are transparently reflected in the data and can be inspected, criticized, and improved in future research.

3.2 Methods

In the following, we present a range of techniques that can be used to detect problems resulting from partial cognacy in phylogenetic reconstruction. Once these problems have been detected, they can be addressed by refining annotations or excluding concepts with high amounts of variation from an analysis.

3.3 Results

We applied the methods described above to the newly compiled data set for Chinese dialect varieties in order to investigate to what degree an extensive number of partial cognates could have an impact on phylogenetic reconstruction analyses. In the following, we will discuss our experiments in detail. We start from our heuristics for the identification of concepts susceptible to high variation due to partial cognacy (Section 3.3.1) and discuss some examples where cognate codings differ, depending on the approach used to make cognacy judgments for entire words from partial cognates. We then carry out a systematic comparison of dialect distances resulting from different coding practices (Section 3.3.2) and conclude by investigating how the coding practice influences the results of phylogenetic reconstruction analyses (Section 3.3.3).

3.3.2 Cognate coding and language distances

Having shown that we can identify quite a few concepts in the Sinitic data in which compounding patterns are so complex that they make the conversion of partial into full cognate sets difficult, we wanted to analyze to what degree this may influence the computation of lexical distances between languages. We therefore computed distance matrices, following classical lexicostatistical methodology (counting shared cognates per meaning slot) for both strictly and loosely converted cognate sets as well as for the two new approaches we introduced in Section 3.2, conversion by common morphemes and conversion by salient morphemes. In order to get a better impression on the theoretical impact which partial cognates can have on lexical distance computation, and the differences between the individual partial cognate conversion schemes, we prepared two distance matrices. In one matrix, only those 59 concepts for which the B-Cubed F-scores would be 0.8 or less were used, and in one matrix all data were used.

In order to compare the two sets of four distance matrices which were the output of this procedure, we used the traditional Mantel test (Mantel, 1967), which calculates the correlation between distance matrices by means of a permutation method, using 999 permutations per run and the Pearson correlation coefficient as our correlation measure. The correlation scores of the Mantel test fall between –1 and 1, with –1 indicating high negative correlation, 1 indicating high positive correlation, and 0 indicating no correlation.

Table 10 shows the result of this comparison. While the correlations are extremely high when taking the full data sets (all 201 concepts) into account, we find more fine-grained differences when inspecting only the subsets. The loose and strict conversion schemes show the highest difference, with a (still high) correlation of 0.71. Our salient morpheme conversion (which is based on the hand-curated assignment of salient as opposed to non-salient morphemes in the data) comes second with respect to its difference from the loose coding scheme and a score of 0.76. The highest correlation between distance matrices can be observed for the salient morpheme scheme and the strict conversion scheme, with a score of 0.96.

Although the correlations between the different coding schemes are all high, even for our worst-case subset, the matrix comparison offers us some clearer insights into the specifics of the different conversion schemes. With the strict and the loose conversion schemes representing two extremes, our two new approaches, automated conversion by common morphemes and hand-curated conversion by salient morphemes, fall between the two extremes, with the salient morpheme conversion—in the way in which it was practiced by us—coming closer to the strict conversion than the common morpheme conversion does.

In order to explore the differences between strictly and loosely converted partial cognates, we visualized the results with the help of heat maps, shown in Fig. 1, where we compare pairwise similarities between the dialects (measured by counting shared cognates) for the strictly and loosely converted partial cognates, using the classification of the seven standard dialect groups by Sagart (2011), later adjusted for subgroups and additional dialect groups by List (2015), as our reference tree. As can be seen from Fig. 1, we have to deal with a lot of reticulation (borrowings or parallel changes due to language contact) in this data set, as reflected in the fact that certain dialects, such as Guìlín (assigned to the Pínghuà group in the source of Liú et al., 2007) or Wēnzhōu (a traditional Wú dialect), show high similarities with the northern dialects (Mandarin and Jìn) in the sample. We also observe considerably low similarity scores between dialects which are traditionally assigned to the same dialect groups, such as Lóudî and Chángshā (Xiāng group). Determining the detailed reasons for these skewed similarities requires a thorough comparison of the individual cognate sets, which would go beyond the scope of this paper. However, that the history of the Chinese dialects is intertwined and contains many reticulate events has been observed in many previous studies (List et al., 2014; Norman, 2003) and should not surprise us too much in this context.

Table 10

Mantel tests of distance matrices derived from a subset of highly divergent concepts (“Subset”) and from considering the full set of data (“Full data set”)

	Subset	Full data set
Loose vs. strict	0.71	0.95
Loose vs. common morpheme	0.85	0.99
Loose vs. salient morpheme	0.76	0.97
Strict vs. common morpheme	0.87	0.96
Strict vs. salient morpheme	0.96	0.98
Common morpheme vs. salient morpheme	0.94	0.99

Mantel tests were calculated from 999 permutations, using the Pearson correlation coefficient as the correlation measure. Significance scores are not provided here, since all permutation tests showed a p-value of less than 0.001, but they are available in the supplementary materials.

View Full Size

Figure 1

Comparing the pairwise similarities in strictly (top) and loosely (bottom) converted partial cognate sets for the dialects in our sample

Citation: Language Dynamics and Change 13, 2 (2023) ; 10.1163/22105832-bja10023

• Download Figure
• Download figure as PowerPoint slide

The differences between the two matrices in Fig. 1 are striking, but difficult to assess from a direct comparison. All in all, and also due to the specific conversion scheme, the loose conversion yields much higher similarity scores than the strict conversion. In Fig. 2, we have tried to visualize these by plotting the differences in the observed distances for strict and loose cognate conversion. We can see that specifically the southern dialects (Mǐn and Yuè), show the largest differences compared to the other dialects in both conversion schemes. The reason for these huge differences, which can reach 20 % in some extreme cases, can be found in the difference between the word structures in northern and southern Chinese dialects. While northern dialects tend to have more multisyllabic words with a complex motivation structure, we find considerably more monosyllabic items in the southern dialects. Since the dialects still employ the same inherited word material, but differ with respect to the compositionality of their words, the strict conversion scheme will increase their divergence, while the loose conversion scheme will increase their similarity.

View Full Size

Figure 2

Differences in shared cognate sets between loosely and strictly converted cognate sets

Citation: Language Dynamics and Change 13, 2 (2023) ; 10.1163/22105832-bja10023

• Download Figure
• Download figure as PowerPoint slide

3.3.3 Partial cognates and language phylogenies

Having analyzed the differences between the distance matrix retrieved from cognate sets derived from partial cognates using different conversion methods, we find that there is a high correlation between all distance matrices when looking at the data set as a whole, while these correlations drop when taking into account only those concepts which we automatically identified as diverse. What remains to be investigated is whether these differences in the distance matrices have a direct impact on the computation of phylogenetic trees. In order to explore this, we took the cognate sets from the 59 highly diverse concepts and generated four Bayesian phylogenies, one for each of the four conversion schemes, following the standard practice of converting cognate sets to binary presence-absence matrices in which language evolution is modeled as a process of cognate gain and cognate loss (Greenhill et al., 2021).

Bayesian phylogenies have become a standard way of inferring phylogenies from lexical data coded for cognate sets. For our analysis, we used the MrBayes software (Ronquist and Huelsenbeck, 2003) and analyzed the data for the four conversion schemes with the help of a fossilized birth-death model (Stadler, 2010), commonly used in Bayesian phylogenetic studies applied to linguistic data (Chang et al., 2015; Sagart et al., 2019). In order to make sure we received comparable results for root ages (also with respect to alternative analyses that have been done on different data sets in the past), we placed the root age between 1,500 to 2,500 years BP, following a uniform distribution. We had the software generate 20,000,000 different trees in two independent runs from which we sampled every 10,000th tree. Low differences between the trees generated in the independent samples indicated that all four analyses reached convergence. Discarding 10 % of the initially generated trees (so-called burn-in), we then reconstructed consensus trees from the remaining 1,800 trees sampled from each of the two runs.

Figure 3 displays the consensus phylogenies reconstructed from the different tree samples. As can be seen from the figure, the tree topologies reconstructed from our four conversion schemes vary quite substantially. Thus, while we find that Hakka (Méixiàn) and Mǐn (Xiàmén and Fúzhōu) form a clade in the strict and the common morpheme conversion, they appear in separate groups in the remaining conversion schemes. While the strict conversion phylogeny provides a scenario in which the more archaic dialect groups of Mǐn, Wú, and Hakka—with the exception of Yuè (Guǎngzhōu), which causes problems in all approaches, probably due to the heavy recent contact with Mandarin—split off first, while more innovative groups are established later, this scenario is less supported by the remaining approaches. With the exception of the loose conversion scheme, in which Chéngdū, a Mandarin dialect, is surprisingly clustered with Xiāng and Wú dialects, all schemes basically recover the traditionally proposed dialect subgroups. The only exception is the Jìn group, represented by Tàiyuán, which is heavily disputed among traditional scholars of Chinese dialectology and classified as a Mandarin dialect in alternative proposals; it appears inside the Mandarin group in all four scenarios.

The scenarios also differ quite substantially with respect to the degree to which the trees are resolved. While we find a clear binary split at the top of the tree only for the strict conversion scheme, we find star-like top-level branchings to different degrees in all other approaches. Here, the loose conversion shows the lowest degree of resolution, failing to resolve eight branches at the top level, followed by the common morpheme conversion with five branches, and the salient morpheme conversion with four branches.

View Full Size

Figure 3

Comparing Bayesian phylogenies (consensus trees) based on our four different conversion schemes: strict conversion (a), loose conversion (b), common morpheme conversion (c), and salient morpheme conversion (d)

Citation: Language Dynamics and Change 13, 2 (2023) ; 10.1163/22105832-bja10023

• Download Figure
• Download figure as PowerPoint slide

Given that we fixed the age of the tree, providing divergence dates conforming to traditional assumptions of Chinese dialect diversification, and given that we did not use any internal calibration points, we cannot learn much from the overall tree ages, which are largely the same in all four approaches. However, internal age estimates show some remarkable differences, specifically for the Wú dialect group, where estimates differ by more than 400 years when comparing the loose conversion estimate of 940 years with the strict conversion estimate of 1,400 years. Similarly, the split of the Mǐn varieties of Fúzhōu and Xiàmén is dated at 1,550 years in the strict conversion, while the three other conversion methods provide estimates of around 1,100 years.

In traditional Chinese historical linguistics, there are different accounts of the overall pattern of Chinese dialect evolution. Norman (2003) assumes that there was a split into three groups, consisting of a southern group comprising Hakka, Mǐn, and Yuè, a northern group consisting of the Mandarin dialects (including Jìn), and an intermediate group consisting of Wú, Xiāng, and Gàn dialects. An alternative scenario, specifically propagated by Karlgren (1954), assumes that the Mǐn dialects split off first, and that the other dialects evolved from a koine that formed around AD 600. Sagart (2011) follows Karlgren (and most Chinese dialectologists) in assuming that the Mǐn dialects split off first, but proposes a more complex diversification scenario, in which the other branches split off step by step, starting from Yuè and Hakka, followed by Wú, Gàn, and Xiāng (see List, 2015, for details on this scenario).

When comparing these scenarios with the phylogenies based on the four conversion schemes, we can see that all four of them diverge from traditional accounts, most likely due to problems in dealing with the impact of undetected borrowings, large-scale convergence in some of the dialect groups, and because the phylogenies were only reconstructed from a small number of concepts susceptible to high variation resulting from lexical compositionality. However, we can also see that the conversion schemes differ regarding the degree to which they diverge from the traditional scenarios. Thus, while the strict conversion scheme conforms in part to the idea of Sagart that Chinese dialect groups split off step by step, the loose conversion scheme proposes a largely star-like diversification of Chinese dialects, in which multiple branches originate from the root at the same time. While the salient morpheme conversion scheme likewise reflects parts of Sagart’s nested scenario in proposing a clade comprising Mandarin, Xiāng, and Gàn (and the highly mixed Pínghuà), the common morpheme comparison only uncovers Mandarin (with Jìn) as a distinct clade, with Gàn as a top-level clade.