Beyond cognacy: historical relations between words and their implication for phylogenetic reconstruction

Abstract

This article investigates the terminology and the processes underlying the fundamental historical relations between words in linguistics (cognacy) and genes in biology (homology). The comparison between linguistics and biology shows that there are major inconsistencies in the analogies drawn between the two research fields and the models applied in phylogenetic reconstruction in linguistics. Cognacy between words is treated as a binary relation which is either present or not. Words, however, can exhibit different degrees of cognacy which go beyond the distinction between orthologous and paralogous genes in biology. The complex nature of cognacy has strong implications for the models used for phylogenetic reconstruction. Instead of modeling lexical evolution as a process of cognate gain and cognate loss, we need to go beyond the cognate relation and develop models which take the degrees of cognacy into account. This opts for the use of evolutionary models which handle multistate characters and allow to define potentially asymmetrical transition tendencies among the character states instead of time-reversible binary state models in phylogenetic approaches. The benefit of multistate models with asymmetric transition tendencies is demonstrated by testing how well different models of lexical change perform in semantic reconstruction on a lexicostatistical dataset of 23 Chinese dialects in a parsimony framework. The results show that the improved models largely outperform the popular gain–loss models. This suggests that improved models of lexical change may have strong consequences for phylogenetic approaches in linguistics.

1. Introduction

Evolutionary biology and historical linguistics both deal with the evolution of objects. Evolutionary biology investigates the evolution of species, morphological characters, and genes, and historical linguistics investigates the evolution of language varieties, grammatical features, and words. In both disciplines, historical relations are an important way to describe the consequences of evolutionary processes. Historical relations are defined for evolving objects which share a common history. The most general historical relation is the relation of common descent. This relation can hold both for lineages and for their characteristics. If the relation concerns the latter, biologists call it homology. In linguistics, this relation is often compared with the relation of cognacy. In contrast to historical relations, we can define various nonhistorical relations between evolving objects. We can compare species for phenotypic similarity and language varieties for typological similarity. We can compare species for the similarity of their habitat, and language varieties for their geographic closeness. Although these similarities can give us hints regarding deeper historical relations, they are neither a necessary nor a sufficient condition for them.

Evolutionary biology has a rich terminological framework describing fundamental historical relations between genes and morphological characters. Discussions regarding the epistemological and ontological aspects of these relations are frequent and fruitful (Jensen 2001; Koonin 2001; Petsko 2001; Sonnhammer and Koonin 2002; Morrison 2015). In historical linguistics, terminological questions regarding historical relations have occasionally been raised in the past (Katičić 1966; Arapov and Xerc 1974), and recent discussions about the cognacy of grammatical features in historical syntax have emerged (Campbell and Harris 2002; Barðdal and Eythórsson 2012; Walkden 2013). In quantitative applications, however, the fundamental historical relations between words, morphemes, or grammatical features are usually assumed to be self-evident, not deserving specific attention. As a result, our traditional terminology dealing with relatedness, inheritance, and descent is often used imprecisely, frequently leading to confusion in quantitative applications. Computational approaches in historical linguistics are often based on software originally designed for bioinformatics. Scholars justify the use of bioinformatics software in linguistics by drawing analogies between historical relations in the two disciplines. Unfortunately, these analogies often ignore the peculiarities of biological evolution and language history. Instead, they offer a simplified mapping between terms in both disciplines and disregard the underlying processes.

In the following, I will try to illustrate the problems in phylogenetic reconstruction in more detail. I will try to show that the models which are currently used to infer phylogenies from linguistic data suffer from a loss of valid information arising from the superficial analogy between homology and cognacy and a simplification of the processes underlying lexical change. Since terminological misunderstandings are the core of the problem, I will first carry out a brief comparison of biological and linguistic terminology on historical relations, pointing to similarities and differences in the two fields (Section 2). By discussing the complexities of lexical change, I will point to further pitfalls that should be avoided when modeling lexical change with biological software (Section 3). I will then propose improvements to the models currently used in computational historical linguistics (Section 4), and illustrate for a small lexical dataset of Chinese dialects how complex historical relations between words can be modeled in computational approaches to phylogenetic reconstruction (Section 5).

2. Terminology for historical relations in biology and linguistics

Scholars have often compared biological and linguistic terminology (Gray 2005; Croft 2008; Pagel 2009; Geisler and List 2013). The analogies that have been made are, however, not necessarily very precise. This becomes especially evident in the analogies drawn between the terms which are used to describe historical relations between evolving objects in both fields. The most popular analogy in this context is that between homology in biology and cognacy in linguistics (Pagel 2009). In the following, I will carry out a detailed comparison between the terminology used in both fields, thereby showing that the analogy between homology and cognacy is essentially misleading.

2.1 Homology

Homology is a fundamental concept in evolutionary biology, designating a ‘relationship of common descent between any entities, without further specification of the evolutionary scenario’ (Koonin 2005: 311). The term was first defined by Richard Owen (1804–92), who distinguished ‘homologues’, as ‘the same organ in different animals under every variety of form and function’ (Owen 1843: 379), from ‘analogues’ as an ‘organ in one animal which has the same function as another part or organ in a different animal’ (Owen 1843: 374). Homology is a very general historical relation between evolving objects. It does not specify the process from which the relation originated. Geneticists distinguish three subtypes of homology based on processes underlying the homology of genes in molecular evolution: orthology, paralogy, and xenology. Orthology refers to ‘genes related via speciation’ (Koonin 2005: 311), paralogy refers to ‘genes related via duplication’ (Koonin 2005: 311), and xenology refers to genes ‘whose history, since their common ancestor, involves an interspecies (horizontal) transfer of the genetic material for at least one of those characters’ (Fitch 2000: 229).

In a paper from 1970, Fitch suggested to distinguish two kinds of homology in molecular evolution: homology as the ‘result of speciation so that the history of the gene reflects the history of the species’ should be called ‘orthology’, and homology as the ‘result of gene duplication so that both copies have descended side by side during the history of an organism’ should be called ‘paralogy’ (Fitch 1970: 113). First evidence that genome evolution does not only involve the mutation of individual genes but also the duplication of genes as a whole was reported in the 1930s (Zhang 2003; Taylor and Raes 2004).

In 1983, Gray suggested to use the term xenology as a third subtype of homology in order to distinguish those cases in which genes are homologous, but neither orthologous nor paralogous, since ‘cells and organisms have acquired foreign genes in the past’ (Gray and Fitch 1983: 64). It is now a well-established fact that prokaryotes (bacteria) may acquire genetic material from ‘their neighborhood or […] environment and incorporate it into their genomes’ (Nelson-Sathi et al. 2013: 166). Lateral gene transfer processes were first detected and described in the 1950s (Freeman 1951). Only 30 years later, however, scholars began to emphasize the importance of lateral gene transfer for microbial evolution (Syvanen 1985). Figure 1 contrasts the three basic processes of speciation, duplication, and lateral transfer with the resulting historical relations in evolutionary biology.

2.2 Cognacy

In historical linguistics, the only relation which is explicitly defined is cognacy (also called cognation). Cognacy usually refers to words related via ‘descent from a common ancestor’ (Trask 2000: 63) and it is strictly distinguished from descent involving lateral transfer (borrowing). The term cognacy itself, however, covers both direct and indirect descent. Hence, German Zahn ‘tooth’ is cognate with English tooth, as is German Kopf ‘head’ with English cup, and German Getränk ‘drink’ with English drink, although the historical processes that shaped the present appearance of these three word pairs are quite different: apart from the sound shape, Zahn and tooth have regularly developed from Proto-Germanic *tan þ (Kroonen 2013: 509f); Kopf and cup both go back to Proto-Germanic *kuppa- ‘vessel’ (Pfeifer 1993; Kluge and Seebold 2002),¹ but the meaning of the German word has changed greatly; Getränk and drink go ultimately back to Proto-Germanic *drinkan ‘to drink’ (Kroonen 2013: 100f), but the German noun was built as a collective (with prefix Ge-) from the nominalized form of the verb (Pfeifer 1993), while the English noun was directly built from the verb. The nominalized form, Proto-Germanic *dranka- is still reflected in German Trank ‘potion’. Thus, of the three examples of cognate words, only the first would qualify as having evolved by direct inheritance. Starostin (2013: 140) suggests to distinguish ‘etymological cognacy’ from ‘lexicostatistical cognacy’, the former denoting words whose ‘forms go back to the same protoform’, and the latter denoting words whose ‘meanings go back to the same meaning in the proto-language as well’. Trask (2000: 234) suggests the term oblique cognacy to label cases in which ‘two or more words in related languages […] continue alternant forms of a single root in the ancestral language’, but this term is rarely used and most of the time linguists simply use the term cognacy without further specifying what they actually mean.

2.3 Beyond homology and cognacy

In an earlier paper (List 2014: 38–46) I abstracted from the processes underlying the historical relations between genes to contrast the biological and the linguistic terminology. In this comparison, I took common descent as the most basic relation, with homology as a direct counterpart. The term ‘common descent’ may be a bit misleading, but what I had in mind by then were all forms of historical relations, including those resulting from lateral transfer. Common descent was further subdivided into direct common descent (corresponding to orthology), indirect common descent (corresponding to paralogy), and common descent involving lateral transfer (corresponding to xenology). I then contrasted the abstract relations and the biological terminology with the terminology currently found in linguistics, thereby pointing to missing slots in the linguistic terminology, for which new terms are proposed. Table 1 illustrates this comparison by contrasting the abstract basic relations with the terminology in biology and linguistics. Relations for which proper terms are missing in linguistics and for which I proposed new terms are colored in red (List 2014: 44). As one can easily see from the table, historical linguistics does not offer direct counterparts for the abstract relations underlying homology, orthology, and xenology in evolutionary biology. Cognacy in historical linguistics is often deemed to be identical with homology in evolutionary biology (Gray 2005; Pagel 2009), but if we follow the comparison, this is only true if one ignores common descent involving lateral transfer, since borrowings are explicitly excluded from the classical definition of cognacy in historical linguistics (Trask 2000: 63).

As we can see from the table, linguistics lacks a proper term for a historical relation between words regardless of whether they are inherited or borrowed (homology in biology, etymological relation according to Table 1). There is also no term denoting the relation between words of which one has been borrowed during its history (xenology in biology). This does not mean, of course, that the relations do not occur in the linguistic domain. Lateral transfer, the process underlying the relation of xenology in molecular biology is also common in language history.² In contrast to a relation between two words which involves lateral transfer, the term borrowing refers to distinct processes involving a donor and a recipient. As an example for such a relation, consider the words German kurz ‘short’ and English short (List 2014: 40). These words are not cognate. German kurz is a borrowing from Latin curtus ‘mutilated’ (Pfeifer 1993), but English short probably goes back to Proto-Indo-European *(s)sker- ‘cut off’ (Rix et al. 2001), and so does Latin curtus (Vaan, 2008). The specific history behind these relations is illustrated in Fig. 2. Since German kurz was borrowed early from Latin, we cannot say that kurz has been borrowed from French court, but we also cannot say that both words are cognate. Yet since both words share a common history, it would be likewise wrong to label them as unrelated, in lack of a proper terminology.

3. Modeling lexical change

In the previous section, I have introduced the basic terminology which biologists and linguists use to denote specific relations between evolving objects. I have then presented an earlier approach of mine (List 2014), where I used the distinctions made in the biological domain in order to introduce new terms for specific historical relations between words. On the first look, the approach seems justified, and the proposed analogies between biological and linguistic relations seem to be fruitful. When looking into the details, however, it becomes clear that important questions are left unanswered. While it is obvious that cognacy in linguistics is not the same as homology in biology, it is less clear how we should understand the idea of direct and indirect cognacy. What exactly is meant to be indirect here? Is it the fact that words differ in meaning, thus being akin to words which are root-cognate but not lexicostatistically cognate, following the distinction of Starostin (2013: 140), or should we instead concentrate on morphological differences, thus following the notion of oblique cognacy proposed by Trask (2000: 234)? And how does the idea of ‘indirect descent’ relate to paralogy and the process of gene duplication in biology? In the following, I will try to show that we need to go beyond my earlier proposal in order to develop a satisfying model of lexical change that can be used for phylogenetic reconstruction.

3.1 Degrees of cognacy

Morrison (2015: 50) points to the relative character of homology in evolutionary biology in emphasizing that evolving objects can exhibit homology at different levels, which may even be independent of each other:

The classic example is the comparison of bird wings and bat wings. These are homologous as forelimbs (structures), which are general throughout the tetrapods, but they are not homologous as wings (functions), because they represent independent modifications of those forelimbs in the ancestors of birds and bats. (Morrison 2015: 50)

We can find similar situations in linguistics: if we consider words for ‘to give’ in the four Romance languages Portuguese, Spanish, Provencal, and French, we can state that both Portuguese dar and Spanish dar are homologous, as are Provencal douna and French donner. The former go back to the Latin word dare ‘to give’, the latter go back to the Latin word dōnāre ‘to gift (give as a present)’. In times when Latin was spoken, both dare and dōnāre were clearly separated words denoting clearly separated concepts and being used in clearly separated contexts. The verb dōnāre itself was derived from Latin dōnum ‘present, gift’. Similar to English where nouns can be easily used as verbs, Latin allowed for specific morphological processes to turn nouns into verbs. What the ancient Romans were not aware of is that Latin dōnum ‘gift’ and Latin dare ‘to give’ themselve go back to a common word form. This was no longer evident in Latin, but it was in Proto-Indo-European, the ancestor of the Latin language. Thus, Latin dare goes back to Proto-Indo-European *deh₃- ‘to give’, and Latin dōnum goes back to Proto-Indo-European * deh₃ - no- ‘that what is given (the gift)’ (Meiser 1998). The word form *deh₃ - no- is a regular derivation from *deh₃-, so on the Indo-European level, both forms are homologous, since one is derived from the other. This means in turn, that Latin dare and dōnum are also homologues, since they are the residual forms of the two homologous words in Proto-Indo-European. And since Latin dōnāre is a regular derivation of dōnum, it means, again, that Latin dare and dōnāre are also homologous, as are the words in the four descendant languages, Portuguese dar, Spanish dar, Provencal douna, and French donner. Depending on the time depth we apply, we will arrive at different homology decisions. The history of the words is depicted in Fig. 3A.

An even more complex example are words like Italian sole, French soleil, Swedish sol, and German Sonne, all meaning ‘sun’. Indo-European scholars assume that the Proto-Indo-European word for sun had a complex, stem-alternating paradigm with two different base forms, one for nominative and accusative case *séh₂$u̯$el-, and one for the oblique cases, *sh$2$én- (Wodtko et al. 2008: 606). Proto-Germanic inherited this paradigm completely (*sōel- versus *sunnōn, Kroonen 2013: 463f), but it was simplified via the process known as analogy in historical linguistics, and the nominative stem was taken as the base form in Latin sōl (Meyer-Lüebke 1911: §8059). In Swedish and German, the complex base form was also simplified, but in different directions, with the Swedish form taking the nominative stem as the basis of analogy, and the German form taking the oblique stem. While Italian sole is the regular reflex of Latin sōl, French soleil goes directly back to Latin sōliculus ‘small sun’, a Latin diminutive of sol (Meyer-Lüebke 1911: §8067). From this perspective, Italian sole is more closely related to Swedish sol than to French soleil, although French and Italian are, of course, much closer genetically related than are Swedish and Italian. The history of the reflexes of the Indo-European word for ‘sun’ is depicted in Fig. 3B.

3.2 Dimensions of lexical change

In a very simple language model, the lexicon of a language can be seen as a bag of words. A word is further defined by two aspects: its form and its meaning. Thus, the French word arbre can be defined by its written form arbre or its phonetic form [ɑɾbɾə], and its meaning ‘tree’. This is reflected in the famous sign model of Ferdinand de Saussure (1857–1913, Saussure 1916), which I have reproduced in Fig. 4A. In order to emphasize the importance of the two aspects, linguists often say that form and meaning of a word are like two sides of the same coin, but we should not forget that a word is only a word if it belongs to a certain language. From the perspective of the German or the English language, for example, the sound chain [ɑɾbɾə] is just meaningless. So instead of two major aspects of a word, we may better talk of three major aspects: form, meaning, and language (Ternes 1987: 22f; List 2014: 15–18). As a result, our bilateral sign model becomes a trilateral one, as illustrated in Fig. 4B.

Gévaudan (2007) distinguishes three dimensions of lexical change: The morphological dimension, the semantic dimension, and the stratic dimension. The morphological dimension points to changes in the form of words which are not due to regular sound change. As an example, consider German Getränk ‘drink’ and its ancestral form Old High German tranc ‘drink’. While the meaning of the word is the same, the German word Getränk is a collective derivation of the Old High German source form (Pfeifer 1993). The derivation process involved prefix Ge-, and the modification of the main vowel. The semantic dimension is illustrated by changes like the one from Proto-Germanic *kuppa- ‘vessel’ to German ‘Kopf’. The stratic dimension refers to changes which involve lexical material outside the historical continuum of a given language (Gévaudan 2007: 141f). In the terminology of Gévaudan (2007: 141f), stratum refers to languages as historical continua, and should not be confused with the way the term is used in sociolinguistics, where it refers to language varieties used in certain layers of a linguistic society (Coseriu 1973; Oesterreicher 2001), but rather in opposition to the term adstratum in historical and areal linguistics (Gévaudan 2007: 141). Usually, changes along the stratic dimension belong to the class of borrowing processes. (Gévaudan 2007: 141–63) argues, however, that processes like onomatopoeia, antonomasy, and folk etymology can also be characterized as processes which involve the stratic dimension of lexical change, since they are based on material which does not stem from the historical continuum of a given language. An example for a simple type of stratic change is English mountain which was borrowed from Old French montaigne ‘mountain’. An example for a more complex type of stratic change is German Maus ‘mouse (for a computer)’ which was not directly transferred from English but rather received a broadened semantic function under the influence of the English word (compare Weinreich (1974: 47–62) and Gévaudan (2007: 143–51) for more details on different types of lexical interference).

Note that these three dimensions of lexical change correspond directly to the three major aspects constituting the linguistic sign: the morphological dimension changes the form of a word, the semantic dimension its meaning, and the stratic dimension its language. Thus, the three dimensions of lexical change, as proposed by Gevaudan find their direct reflection in the major dimensions along which words can vary.

3.3 27 Shades of cognacy

When looking at the different historical relations from the perspective of the three dimensions of lexical change, it becomes clear that the new terms I proposed earlier (List 2014) do not necessarily solve our problem of reflecting the different aspects of lexical change and lexical variation adequately. Although it seems justified to point to the difference between cognacy in linguistics and homology in biology, it proposes a problematic analogy between paralogy and indirect cognacy without further specifying how indirect cognacy should be defined in the end. When investigating the different uses of the term ‘cognacy’, for example, it becomes obvious that the differences result from controlling for one or more of the three dimensions of lexical change proposed by Gévaudan (2007).³ The notion of cognacy of a classical Indo-Europeanist, for example, controls the stratic dimension by requiring stratic continuity (no borrowing), but at the same time it is indifferent regarding the other two dimensions. This is what Starostin (2013: 140) called ‘etymological cognacy’. Cognacy à la Swadesh (especially Swadesh 1952,1955), as we know it from lexicostatistics (Swadesh 1952, 1955) and its modern derivations (Gray and Atkinson 2003), is indifferent regarding morphological continuity, but controls the semantic and the stratic dimensions by only considering words that have the same meaning and have not been borrowed. This is what Starostin, (2013: 140) called ‘lexicostatistical cognacy’.

‘Traditional cognacy’ and ‘cognacy à la Swadesh’, however, are but two ways to control for the three dimensions of lexical variation, and one can easily think of more perspectives on historical relations between words, including the terminology that is used in evolutionary biology. In Table 2, I have attempted to illustrate in which way the different terms, including the biological terms of homology, orthology, and xenology, cover processes by controlling each for one or more of the three dimensions of lexical change (with + indicating that continuity is required, − indicating that change is required, and +/− indicating indifference). Note that paralogy was not included in the comparison, since the process of gene duplication is a very specific event that probably has no fruitful analogy in historical linguistics. Contrasting the different dimensions of lexical change with the terminology used to refer to different relations between words shows the arbitrariness of the traditional linguistic terminology. Why do we only cover two out of $3*3*3=27$ different possible types? Why do we only control by requiring continuity, not change? It also shows the fundamental difference between change processes in linguistics and biology.

4. Models of lexical change in phylogenetic reconstruction

In the previous sections, I have tried to show that not only the terminology that we use to denote historical relations between evolving entities in linguistics and biology shows some important differences, but also that the processes underlying lexical change in language history are very particular, involving three major dimensions of lexical variation which themselves can be further subdivided into a multitude of minor process types.⁴ In the following, I will try to illustrate how our models can be modified in order to account for more complex historical relations between words.

4.1 Gain loss models and morphological variation

The majority of automatic methods for phylogenetic reconstruction in historical linguistics employ lexical data to infer language phylogenies. When employing these methods, it is important to specify a model of lexical change that the algorithms can use to infer the trees or the networks that fit the data best. Most datasets employ a lexicostatistical scheme of data-coding (Dyen et al. 1992; Ringe et al. 2002; Greenhill et al. 2008; Bouckaert et al. 2012; Greenhill 2015). This means, that they are based on concept lists of 100 and more items which are translated into the languages under investigation. By comparing all translations in each concept slot with each other, linguists then annotate which words are cognate. The notion of cognacy that is underlying these databases is usually the notion of ‘cognacy à la Swadesh’ in Table 2, that is, annotators try to filter out borrowings, consider only semantically identical items, and do not necessarily regard morphological variation.

The methods which are then used to analyze the data, be they based on probabilistic approaches (Felsenstein 1981; Huelsenbeck et al. 2001), or parsimony (Fitch 1971; Sankoff 1975), are almost exclusively based on gain–loss models of lexical change (Pagel 2009). They reduce the change of phylogenetic characters to processes of gain and loss and essentially assume that during evolution a language can either gain a new word or loose an existing one. In these models, each phylogenetic character has only two states, presence, or absence, and presence–absence matrices of cognate sets are fed to the algorithms in order to infer language phylogenies. Presence–absence matrices are retrieved from the original data by breaking up the semantic slots into sets of cognate words, and listing for each language whether it has a word belonging to the respective cognate set or not (Atkinson and Gray 2006). This way of data preparation and encoding is further illustrated in Table 3.

The binary coding practice has strong consequences, since it is vulnerable to historical word relations with variation along the semantic and the morphological dimension. First, the general procedure by which lexicostatistical data is binary encoded and concepts are split into several independent characters creates dependencies which cannot be observed by the algorithms. It deprives the analysis of the essential criterion for gain and loss, since presence and absence are defined with respect to meaning identity. Gain and loss need to be essentially interpreted as gain and loss with respect to a certain concept slot, not with respect to the entire language. The loss of a word means that the word is no longer used to express a certain meaning, and the gain of a word implies that a new word is used to express a certain meaning. Yet since meaning is discarded by the binarisation procedure (see Table 3), the models are given no clue to handle instances of parallel semantic shift. A more realistic gain–loss analysis should include a larger sample of words and annotate cognates regardless of differences in meaning (Michael et al. 2015).

Second, the lexicostatistical coding practice is vulnerable with respect to morphological change, since morphological variation is deliberately ignored when assigning words to cognate sets. This was not the case in the early days of lexicostatistics. Hattori (1961), for example, distinguished clearly between true ‘orthologues’ and morphologically derived words. Recalling the example of Italian dare and French donner given in Fig. 3, it is clear that we can annotate the words quite differently, depending not only on the “shade” of cognacy we choose, but also on the desired depth of analysis. In current practice, words like dare and donner are usually assigned to the same cognate set, and their morphological differences are ignored.⁵ When annotating the words, however, we should ask ourselves which kind of annotation would be the best for the underlying model that we use. From this perspective, we would do best in coding Italian dare and French donner as being not cognate, since by the time that donner replaced earlier dare in the ancestor of French, the word dare was lost with respect to the meaning ‘to give’, and the word donner was gained.

The problem of morphological variation in lexicostatistical datasets becomes even more evident when looking at more specific processes of morphological change like compounding. While compounding is less characteristic for the Indo-European language familiy (at least as far as the stable parts of the lexicon are concerned), it plays an important role in the Sino-Tibetan language family (Matisoff 2000: 341f; Chung et al. 2014; List 2015: 56–58). In the Chinese dialects, for example, the majority of words is only indirectly related, as illustrated in Table 4 where the words for ‘moon’ in four Chinese dialects share the same base morpheme, but differ regarding the further parts of their compounds. When investigating these patterns, we can immediately infer processes of lexical change that link these patterns. Fúzhōu [ŋuoʔ⁵] 月, for example, reflects the oldest stage in which Chinese was still predominantly monosyllabic. Měixiàn [ŋiat⁵ kuoŋ⁴⁴] 月光 reflects a younger stage in which bisyllabic structures were gaining ground, and Wēnzhōu [ɲjy²¹kuɔ³⁵vai¹³] 月光佛 reflects an even later stage, since it builds on the form in Měixiàn, adding a suffix that marks nominalization (compare Wēnzhōu [ɲji²¹dʏu³⁵vai¹³] 日頭佛 ‘sun’).⁶ In the ‘classical’ lexicostatistical view of cognacy and the ‘classical’ models of word gain and word loss, these processes are all ignored, although they may bear important phylogenetic information. One would either label all four words as cognate, since they share the same base morpheme (Satterthwaite-Phillips 2011: 95–103), or label them all as not being cognate, since their parts to not match completely (Ben Hamed and Wang 2006; Gates 2012: 51). If we want to model the evolution of the four words for ‘moon’ in the four dialects realistically, neither of the two encoding practices will be of use. In both cases, all phylogenetic signal will be lost and the analysis cannot tell us how the words really developed (see Fig. 5A and B).

4.2 From binary to multistate models

In principle, phylogenetic methods can handle semantic and morphological variation sufficiently. All we need to do is to switch from binary gain–loss models to multistate models. In a binary state model each character can only be present or absent in a given language, like the cognate set 1 in Table 3, for example, which is present in German and English but absent in Italian and Spanish. In a multistate model, a character cannot only be present or absent, but it can also vary among languages and occur in different shapes. Instead of labeling French donner and Italian dar either as exclusively cognate or as exlusively noncognate, we could assign both words to the same character but assign them different states. In this way, we could handle both variation along the semantic and the morphological dimension of lexical change. If we can further determine how likely it is for the character to switch from one particular state to another, we can force our algorithms to prefer certain transitions and to ignore others. In the case of the Chinese words for ‘moon’ in Table 4, for example, we already saw that Měixiàn [ŋiat⁵ kuoŋ⁴⁴] 月光 is particularly close to Wēnzhōu [ɲjy²¹kuɔ³⁵vai¹³] 月光佛, since the latter was only extended by one suffix. When comparing the Wēnzhōu form with the form [ŋuoʔ⁵] 月 in Fúzhōu, we can further easily say that the transition from the Fúzhōu form to the Měixiàn form should be easier to accomplish than the direct transition to the Wēnzhōu form. If we further know that the process we are dealing with has strong unidirectional tendencies, as it is the case for many processes of sound change and grammaticalization (Haspelmath 2004), but also in inflectional morphology (Wurzel 1985), and potentially even in analogy (Jacques 2016), we can model this by using irreversible models in our analyses (Huelsenbeck et al. 2002; Bohl and Lancaster 2003).

In a parsimony framework of phylogenetic reconstruction (Fitch 1971; Sankoff 1975), the difficulty of switching between the different states of a character is handled by defining specific weights for character state transitions. If we further know that the process we are dealing with has strong unidirectional tendencies, we can model this by assigning asymmetric weights for the transition preferences between the states of a character. The differences between gain–loss models and multistate models allowing for asymmetric transition preferences in a parsimony framework are exemplified in Fig. 5, but multistates and asymmetric transition tendencies can essentially also be handled in probabilistic frameworks.

5. Using improved models to study Chinese dialect history

In order to illustrate the benefits of improved models for lexical change, I have prepared a small experiment on Chinese dialect history. In this experiment, I test how well different models of lexical change with varying degrees of complexity perform on the task of semantic reconstruction. In classical historical linguistics, semantic reconstruction seeks to infer the original meaning of a set of cognate words (Fox 1995: 115–6). The experiment I designed follows lexicostatistical approaches in which semantic reconstruction seeks to identify the word form which was used to express a certain concept in an ancestral language (Kassian et al. 2015: 304–6). In this context, semantic reconstruction can be treated as a specific type of ancestral state reconstruction (Pagel 1999) applied to lexicostatistical data. The starting point is a lexicostatistical wordlist, consisting of a list of concepts which are translated into a set of language varieties. Concepts comprise phylogenetic characters, and the counterparts of the concepts in the respective language varieties reflect different states of the characters. Semantic reconstruction starts from a reference phylogeny (a phylogenetic tree) and tries to infer which character state was present at the root. Chinese is attested through its contemporary dialects, whose diversity is at least comparable to that of the Romance languages (Wang 1997), but also in ancient texts predating the diversification of the modern dialect varieties by several hundred years.⁷ Therefore, in the majority of cases, there is independent evidence regarding the words which were originally used to express a given concept. For this reason, Chinese is an ideal candidate to test the performance of different models of lexical change.

5.1 Materials

The data for the experiment were originally compiled for the study of Ben Hamed and Wang (2006). It comprises 200 concepts translated into 23 Chinese dialect varieties. The concept list is largely identical with the list of 200 items proposed by Swadesh (1952).⁸ In the data, partial cognate relations are annotated by listing the ‘etymological character’ for each morpheme of a word (běnzì 本字, Branner 2000: 35). This information is regarded as problematic by some Chinese dialectologists (Branner 2000), since it is not necessarily clear how consistently the morphemes in dialect words are identified. Datasets like the one by Ben Hamed and Wang (2006) are, nevertheless, a useful starting point for experiments on morphological processes in lexical evolution, especially since other collections which list information on partial cognacy in such great detail are not available. For most of the cases, however, we can assume that the assignments are correct. In an earlier study (List 2015), I used the data by Ben Hamed and Wang (2006) and converted it into a machine-readable text format, which I used for this experiment. All data were thoroughly checked and refined, since the partial cognate assignments were not the primary target of my earlier study and therefore only inconsistently converted into text format.

Ben Hamed and Wang (2006) also give the ancestral forms for the concepts in Old Chinese. Since Old Chinese is supposed to be the ancestor of all dialect varieties in the sample, the data can be used as a ‘gold standard’ to test the accuracy of ancestral state reconstruction methods. Since processes of lexical evolution are quite different for nouns and verbs, with compounding and partial cognacy occurring almost exclusively on nouns, only nouns were considered for this study. Of the 85 concepts denoting nouns in the sample, 28 were excluded. Either the reflexes were all different from the Old Chinese forms and it would be impossible to reconstruct them, or the reflexes were all identical with the Old Chinese form, and reconstruction would be no challenge at all. The 57 forms considered for the experiment are listed in Table 5 along with the supposed ancestral forms in Old Chinese.

Ancestral state reconstruction requires a reference phylogeny as input. Here I build on an earlier approach (List 2015) where I compared reference phylogenies for three independent hypotheses on Chinese dialect history, namely Laurent Sagart’s Arbre des Dialectes Chinois (Sagart 2011), the Hànyǔ Fāngyán Shùxíngtú 漢語方言樹形圖 (‘Tree chart of Chinese dialects’) by Yóu Rǔjié 游汝傑 (Yóu 1992: 91–106), and Jerry Norman’s Southern Chinese Hypothesis (Norman 1988: 210–4). These reference phylogenies differ regarding the subgrouping of the seven major dialect groups of Chinese and are based on competing criteria for subgrouping (see List 2015: 36f for details).

5.2 Methods

The experiment employs a parsimony framework for character transitions (Nunn 2011: 59–63). Parsimony was used for reasons of simplicity and data sparseness. Parsimony applications can be easily implemented from scratch, while there are no available ready-to-use implementations of probabilistic approaches which handle asymmetric transitions between multiple character states. Given the sparseness of the data available for testing, it is also not clear whether probabilistic applications would converge. Four different models of varying complexity were defined for the experiment:

• BINARY: Character states which are not completely identical in their compound structure are split into sets of binary characters following the classical procedure described in Atkinson and Gray (2006). Character transitions are modeled as a gain–loss process.

• FITCH (multistate): Lexical evolution is modeled as a process of character transitions with equal weights, following the classical model by Fitch (1971).

• SANKOFF (multistate, weighted): Lexical evolution is modeled as a process of character transitions with unequal weights, following the classical model by Sankoff (1975).

• DWST (‘directed weighted state-transitions’, multistate, weighted, directed): Lexical evolution is modeled as a process of character transitions with unequal weights and in dependence of the direction of the transition.

The BINARY and the FITCH model are straightforward in their implementation. The BINARY model only handles gains and losses with losses being favored over gains. The parsimony weight for gain events was set to 2, and the penalty for loss events was set to 1, since these penalties yielded the most plausible scenarios in earlier experiments on the data (List 2015). The FITCH model gives equal weights to transitions between all states. In the case of SANKOFF and DWST, transitions are weighted differently depending on the character states. Since we lack exhaustive linguistic accounts on processes of compounding in the Chinese dialects, a very simple approach for the computation of the weights was employed. In a first step, the morpheme representation of two words, which is given in Chinese character readings, with identical characters representing cognate morphemes, was aligned using the Needleman–Wunsch algorithm (Needleman and Wunsch 1970). In a second step, it was counted in how many positions the aligned sequences differ. This distance, commonly known as the Hamming distance (Hamming 1950), was further refined by counting substitutions (those instances where two different morphemes are aligned) twice, and insertions and deletions (those instances where a morpheme was aligned with a gap symbol or vice versa) only once. In contrast to the SANKOFF model, the computation of weights for the DWST model only reduces the weights for insertions (a gap aligned with a morpheme), but not for deletions. This transition schema accounts for the tendency of disyllabification in the history of Chinese, during which most of the monosyllabic words in the Chinese dialects were replaced by bisyllabic compounds. Figure 6 gives examples for the differences in the transition penalties of the multistate-models (FITCH, SANKOFF, and DWST) and the calculation of the transition penalties for the SANKOFF and the DWST model. It is beyond doubt that the models could be further refined, and potentially also trained. For the purpose of the experiment, however, it is advisable to keep the models as abstract as possible. This guarantees that we do not overly fit the models to the data, and it also makes it easier to determine the major factors that determine differences in their performances.

The models and the code to optimize the parsimony score were implemented in Python. The code requires the LingPy software package for quantitative tasks in historical linguistics (List and Moran 2013) to calculate the alignments between the characters states and the transition probabilites. The source code along with the data, the results, and further instructions on how to replicate all analyses presented in this article are provided as supplementary data.

5.3 Results

With four different models and three different reference phylogenies, 12 different tests needed to be carried out. In order to evaluate the quality of semantic reconstruction, a simple approach was used. In this approach, one counts the amount of hits and fails. For each concept, all ancestral forms proposed by a given test were considered and compared with the known forms in the ‘gold standard’. If only one form was proposed, this form can either be a hit or a fail, that is, it can either be identical with the form in the gold standard, or not. If multiple forms are proposed by an algorithm, the score is divided among hits and fails, following the proportion of correctly and incorrectly proposed ancestral forms. If, for example, two forms are proposed of which only one is correct, this would be scored as a 50% hit and a 50% fail. The results were evaluated separately for each meaning slot and then averaged across all 57 concepts in the sample.

Table 6 shows the detailed results for all 12 different analyses, including the overall parsimony scores obtained. The DWST model performs best in all respects, regardless of the reference phylogeny. The SANKOFF model outperforms the remaining two models, but only when applied to the Arbre reference phylogeny, it comes close to the high scores of the DWST model. Whether the BINARY or the FITCH model performs better is hard to say, given that the differences are minimal on average, and both models seem to rely heavily on the reference phylogeny. What is remarkable is that the DWST model does not only show the highest scores, but also a high resistency regarding the underlying reference phylogeny. According to the analysis by List (2015), the Arbre gives a more realistic picture of Chinese dialect history. This is reflected by the highly improved scores of all models (except from DWST) for the Arbre phylogeny as opposed to Shùxíngtú and Southern Chinese.

Parsimony approaches may yield multiple solutions which are all optimal with respect to the transition penalties defined in a model. Depending on the character and tree topology, the amount of optimal scenarios may vary greatly. In the FITCH analyses, for example, the number of possible scenarios for all characters ranges from 1 (for ‘ash’) to 4 797 (for ‘night’). As expected, the number of possible scenarios decreases when increasing the complexity of the models. This is shown in Table 7 where the proposed proto-forms and the number of possible scenarios for the analysis of three concepts using the three multistate models are displayed. The table shows clearly that complex models reduce the uncertainty with respect to alternative scenarios.

Figure 7 shows one of four possible scenarios for the development of reflexes of ‘moon’ inferred by the DWST model for the Arbre reference phylogeny. The scenario proposes a pattern in which the word form yuè 月 ‘moon’ was replaced by the compound yuèguāng 月光 ‘moon-light’ in all dialects except from the Mǐn subgroup. While this may well reflect a realistic scenario, we also find homoplastic (reoccurring) transitions, especially from yuèguāng 月光 to yuèliàng 月亮 ‘moon-shine’ in the Wú subgroup. Homoplasy may point to lateral transfer events (List et al. 2014, Dagan and Martin 2007), but our knowledge regarding lexical evolution during the history of the Chinese dialects is still very limited. It is extremely difficult to tell with certainty whether the common reflexes of yuèliàng 月亮 in the Běijīng-Xiāng and the Wú subgroup reflect independent parallel developments or areal influence.

6. Conclusion

In this article, I have pointed to problems in the models used for phylogenetic reconstruction in linguistics resulting from a superficial treatment of historical relations between words. Cognacy is not a binary relation which is either present or not. Instead, we can distinguish different subtypes of cognacy, just as biologists can identify specific types of homology between genes. In an earlier paper, I proposed to compare the biological subtypes of homology (orthology, paralogy, xenology) directly with potential subtypes of historical word relations in linguistics (List 2014), but by concentrating on the major dimensions of lexical change proposed by Gévaudan (2007), namely morphological, semantic, and stratic change, I have shown that we can even go beyond the biological terminology and set up fine-grained schemas for historical relations in linguistics.

Which notion of cognacy we use for phylogenetic reconstruction crucially depends on the data we have at hand and the algorithms we intend to employ. I have shown that the inconsistencies in the treatment of historical relations between words have a direct impact on the way cognates are coded and data are analyzed in phylogenetic approaches. This was illustrated in detail for historical relations involving morphological change, especially compounding. If compounding is frequent and characteristic for a given language family, phylogenetic approaches which model lexical change merely as a process of cognate gain and cognate loss are inadequate and unrealistic. In order to take the different degrees of cognacy into account, I proposed to employ multistate instead of binary state models, and to further allow for potentially asymmetric transition tendencies among character states. The benefits of these models were demonstrated in a small experiment on semantic reconstruction applied to a lexicostatistical dataset of 23 Chinese dialect varieties. The results of this experiment strongly suggest that multistate models with asymmetric transition tendencies are superior to binary state models. What I have presented is, however, but a small step toward improved models of lexical change. More experiments including more language families need to be carried out. Instead of ancestral state reconstruction, we need to test the potential of multistate models for phylogenetic reconstruction in general. Probabilistic models, be they based on Maximum Likelihood (Felsenstein 1981) or Bayesian inference (Huelsenbeck et al. 2001), may prove really useful in this regard. In parsimony, we need to provide exact models for the transition between characters, and we always run the danger of overfitting our step matrices on a given dataset. Probabilistic models can help to estimate transition probabilities and could thus even provide new insights which go beyond cognacy and help us to detect major trends in lexical evolution, including morphological, semantic, and stratic change. In order to allow for these improved models of lexical change, however, we need to rethink the way we handle cognacy in our databases and start being more explicit in our annotations.

Acknowledgements

I thank three anonymous reviewers for challenging critics and helpful advice, and Gerhard Jäger for helpful comments made on an earlier version of this article. I am very grateful to Hans Geisler, who originally pointed me to many of the examples on lexical change which were treated in this article, to David Morrison, for important comments on the article and numerous fruitful discussions on questions of homology and cognacy, and to Laurent Sagart, who shared his ideas and his profound knowledge of Chinese dialect classification with me.

Funding

This research was supported by the DFG research fellowship grant ‘Vertical and lateral aspects of Chinese dialect history’ (Grant No. 261553824), which is gratefully acknowledged.

Footnotes

References