Introduction

The so-called Altaic hypothesis suggests common ancestry for several universally accepted language families spoken across Eurasia, namely the Turkic, Mongolic, Tungusic, Korean and Japonic families (see Fig. 1 for the map). The hypothesis has two common versions. The ‘narrow’ one claims that Turkic, Mongolic and Tungusic are genealogically related to each other and form a distinct clade. The ‘broad’ version expands this clade to also include Korean and Japonic. The resulting clade is traditionally labelled as the Altaic macrofamily, although certain researchers use the term ‘Altaic macrofamily’ specifically for the [Turkic, Mongolic, Tungusic] clade, whereas the same plus Korean and Japonic can be referred to as the ‘Macro-Altaic’ or, using a recently introduced term, ‘Transeurasian macrofamily’. Below we prefer to keep the term ‘Altaic’ for the Altaic macrofamily in the broad sense, assigning the special term ‘Nuclear Altaic’ to the [Turkic, Mongolic, Tungusic] clade.

Figure 1. Modern distribution of the five language families that constitute the Altaic (a.k.a. Transeurasian) macrofamily.

A very detailed and extensive overview of the Altaic hypothesis, its history and the current state of art is offered in Blažek (Reference Blažek2019), with the recent handbook (Robbeets & Savelyev, Reference Robbeets and Savelyev2020) nicely complementing it. As of today, the only attempt at a systematic reconstruction of Proto-Altaic (= Macro-Altaic) phonology and vocabulary has been offered in Starostin, Dybo, & Mudrak (Reference Starostin, Dybo and Mudrak2003); however, it has been severely criticized for various perceived methodological flaws by both opponents of the Altaic hypothesis (e.g. Vovin, Reference Vovin2005) and its supporters (e.g. Robbeets, Reference Robbeets2005, offering a major revision of the reconstructed phonology and rejecting a large number of proposed Altaic etymologies).

In general, most experts agree that lexical similarities between the ‘narrow’ Altaic taxa – Turkic, Mongolic and Tungusic – are too numerous to be accidental, but the proposed historical scenarios differ: some scholars argue in favour of genealogical relationship, while others attribute the observed lexical matches to several chronological layers of (prehistoric) contacts. The same arguments are applicable to the relations between Korean and Japonic, while their hypothetical connections to Turkic, Mongolic and Tungusic are less obvious and actually raise the issue of potential chance resemblances.

A complicated, but historically realistic, ‘hybrid scenario’ (usually adopted by ‘Altaicists’, albeit to highly differing degrees) could imply that an original deep-level genealogical relationship between Turkic, Mongolic, Tungusic, Korean and Japonic may have later become overlapped and partially obscured by various contacts between the already separated branches, varying in intensity and continuance.

As far as we know, the only previously published statistical tests for Altaic relatedness are Turchin, Peiros and Gell-Mann (Reference Turchin, Peiros and Gell-Mann2010) and Ceolin (Reference Ceolin2019). The first of these studies (Turchin et al., Reference Turchin, Peiros and Gell-Mann2010) applies a bootstrap test to the data, consisting of 100-item wordlists encoded for analysis in accordance with the so-called ‘consonant class’ method (in which each form is reduced to a ‘consonantal type’ based on its two first consonants). The languages involved include Proto-Turkic, Proto-Mongolic, Proto-Tungusic and Proto-Japonic (no Korean). However, the analysis suffers from several shortcomings, such as:

1. (1) There is no methodology of semantic (onomasiological) reconstruction. Protoforms and their meanings were taken from Starostin et al. (Reference Starostin, Dybo and Mudrak2003) and other databases of the Tower of Babel Project, although precise semantic reconstruction was never a goal of the project and individual databases can seriously vary in quality in respect to semantic accuracy.

2. (2) The accepted allocation of consonants to classes seems unreasonably coarse-grained: only nine consonant classes are introduced and some decisions look controversial, e.g. r-type phonemes are united with t-type phonemes as a single class, even though it is typologically more natural to either join r- and l-type phonemes in a single class or (as we prefer) assign a distinct class for r-type sounds.

3. (3) No multiple testing correction was applied to the results of individual bilateral comparisons.

Despite these points, Turchin et al.'s experiment is somewhat similar to the one proposed in the present paper.

A recent paper (Ceolin, Reference Ceolin2019) represents a completely different paradigm. The dataset includes 168-item wordlists from three standardized languages, representative of the three Nuclear Altaic families: Literary Turkish, Standard Khalkha Mongolian and Written Manchu. The author attempts to revive some archaic techniques of machine wordlist comparison previously proposed by other scholars (Ringe, Kessler and Lehtonen), e.g. for Indo-European–Uralic comparison. The flaws of Ceolin's experiment appear to be exactly the same as those of the preceding IE-Uralic attempts overviewed and criticized in Baxter (Reference Baxter, Salmons and Joseph1998) and Kassian, Starostin and Zhivlov (Reference Kassian, Starostin and Zhivlov2015: 302–304). The following are the most important objections:

1. (1) The choice of automatic algorithms applied to the data is not substantiated by the author; moreover, it explicitly contradicts the techniques normally used in comparative-historical linguistics. The main problem is that only the initial root consonant is taken into account. This is the same as if we were to investigate the language relationship between English and German by focusing on such lexical pairs with similar initial consonants as belly/Bauch, bird/Vogel, woman/Frau. In actual practice, historical linguists always attempt to compare whole roots rather than arbitrarily selected parts of roots. Because of this, and since the most common root pattern in the world's languages is CVC(V) (mono-consonantal structures CV and VC may formally be regarded as variants with a zero-class consonant in the first or second position – CH and HC respectively), the automated bi-consonantal comparison between CC-forms as proposed by Dolgopolsky (Reference Dolgopolsky1964, Reference Dolgopolsky, Shevoroshkin and Markey1986), then adapted by Turchin et al. (Reference Turchin, Peiros and Gell-Mann2010) and then by Kassian et al. (Reference Kassian, Starostin and Zhivlov2015) and employed for the present study, is more appropriate for machine analysis when it comes to minimizing false positive matches.

2. (2) As a result, the applied automatic algorithms reveal a substantial number of incorrect etymological pairs and improbable sound correspondences, e.g. within the Indo-European test group, Ceolin's (Reference Ceolin2019: 313) machine comparison between such recognizably related lects as Modern Italian and Standard Hindi recognizes the phantom correspondence ‘Italian l-/Hindi c-’ deduced from the following ‘etymologies’: It. largo/Hi. caura ‘wide’, It. liscio/Hi. cikna ‘smooth’, It. luna/Hi. cand ‘moon’ (not a single one of these pairs represents true cognates). Similarly odd sound correspondences can be found in Ceolin's (Reference Ceolin2019: 318) Altaic comparison. Thus, Turkish k-/Manchu f- involves such ‘etymologies’ as:

   • • Turkish kɨsa ‘short’ (< Proto-Turkic *kɨs-ka ‘short’ from Proto-Turkic *kɨs- ‘to press, squeeze’; Dybo, Reference Dybo2013: 542) = Manchu foxolon ‘id.’ (< Proto-Tungusic *poko-lo- ‘short(?)’; Starostin et al., Reference Starostin, Dybo and Mudrak2003: 1100; Tsintsius, Reference Tsintsius1977: 331);

   • • Turkish kök ‘root’ (< Proto-Turkic *kök ‘a k. of root’; Dybo, Reference Dybo2013: 435) = Manchu fulexe ‘id.’ (Tsintsius, Reference Tsintsius1977: 302; no comparanda in Tungusic, but the Proto-Tungusic form is expected to be *pule-; Starostin et al., Reference Starostin, Dybo and Mudrak2003: 1182), and so on. Likewise, Khalkha nV-/Manchu V-involves such ‘etymologies’ as:

   • • Khalkha neg ‘one’ (< Proto-Mongolic *nike-n ~ *nige-n; Gruntov & Mazo, Reference Gruntov and Mazo2015: 235; Nugteren, Reference Nugteren2011: 460) = Manchu emu ‘id.’ (< Proto-Tungusic *emu; Starostin et al., Reference Starostin, Dybo and Mudrak2003: 505; Tsintsius, Reference Tsintsius1977: 270–272);

   • • Khalkha nawč ‘leaf’ (< Proto-Mongolic *labči-n ~ *nabči-n; Gruntov & Mazo, Reference Gruntov and Mazo2015: 231; Nugteren, Reference Nugteren2011: 450) = Manchu abdaxa ‘id.’ (< Proto-Tungusic *xabda-nsa; Starostin et al., Reference Starostin, Dybo and Mudrak2003: 764; Tsintsius, Reference Tsintsius1975: 5), and so on.

   We are not aware of any version of Proto-Altaic reconstruction or of any proposed scenario of lexical contact that would imply that the forms within the aforementioned pairs could be related to each other. Given that these are but a few examples quoted from a large group of similarly false positives, it is not clear whether the obtained results could be considered trustworthy even from a purely theoretical stance.

3. (3) The wordlists were manually designed in order to reinforce the signal, raising the issue of personal bias. Thus, the concept ‘man (male)’ was intentionally included in the list (Ceolin, Reference Ceolin2019: 308) for the sake of the match between Turkish er ‘man (male)’ and Khalkha er ‘id.’, but at the same time the mismatched Manchu xaxa ‘id.’ was deleted from the dataset owing to its allegedly onomatopoeic nature (it is actually very much an open question how justified it is to automatically consider any C ₁V ₁C ₁V ₁-stem encountered in any of the world's languages as onomatopoeia).

4. (4) Compiled wordlists suffer from inaccurate lexicographic work. Thus, e.g. within the Khalkha 110-item wordlist (a subset of the whole Khalkha 168-item wordlist used by Ceolin), the following errors have been detected, the majority of which directly affect Ceolin's test results:

   • • ‘earth’ – Khalkha towrog actually means ‘dust’ and must be excluded. The basic Khalkha term šoroː ‘earth (soil)’ is not listed at all. This is crucial for the experiment since Khalkha towrog matches Turkish toprak ‘earth’ in terms of initial consonants;

   • • ‘to hear’ – the rare Khalkha form duːla- specifically means ‘to heed, hearken’ and must be excluded. This is crucial for the experiment since Khalkha duːla- matches Turkish duymak, Manchu donjimbi ‘to hear’ in terms of initial consonants;

   • • ‘horn’ – Khalkha büreː actually means ‘horn (music instrument)’ and must be excluded. This is crucial for the experiment since Khalkha büreː matches Turkish boynuz ‘horn’ in terms of initial consonants;

   • • ‘root’ – the quoted Khalkha word ug ‘root’ is a marginal rather than basic term;

   • • ‘short’ – Khalkha axar ~ oxor ‘short’ exists, but it is a marginal term with limited lexical compatibility, so it must be excluded;

   • • ‘to sleep’ – the quoted Khalkha verb noyrso- ‘to sleep’ is a marginal rather than basic term;

   • • ‘small’ – Khalkha ǯaːl actually means ‘baby; small (of age)’, ǯiǯig actually means ‘minute, petty (Russian melkij)’ rather than generic ‘small’, whereas the basic Khalkha term očüːxen ‘small (of physical objects)’ is not listed at all. This is crucial for the experiment since Khalkha očüːxen would match Manchu ajige ‘small’ in terms of initial consonants;

   • • ‘year’ – quoted Khalkha ǯil means ‘annual cycle, calendar year’. A more basic term is on ‘year (as a time period)’. At the very least, both should be taken as synonyms. This is crucial for the experiment since Khalkha on would match Manchu aniya ‘year’ in terms of initial consonants.

Most importantly, it may be a priori questioned whether a permutation test conducted on three modern (or historically attested) literary languages, at least two of which (Khalkha and Manchu) have been in a state of intense linguistic contact with each other over the past several centuries, could yield a proper positive signal for a genetic relationship between their ancestors that is generally understood (by those who support the Altaic hypothesis) to date to a period preceding Indo-European. In our opinion, to stand at least a reasonable chance of success, such a test should be conducted not on a small number of living descendants, all of which have undergone significant cognate loss over millennia, but on the reconstructed ancestral stages of these languages – provided, of course, that the latter have been produced based on a more or less rigorous application of the classic comparative method.

The goal of our present project is to estimate the degree of statistical support for the Altaic hypothesis with the help of an algorithm which could, at least to a certain extent, emulate the research of an actual comparative linguist. It is the combination of such a research procedure, well tested over the past couple of centuries, with the formal edge provided by machine algorithms that could, in our opinion, provide the best possible outcome for any controversial hypotheses of distant linguistic relationship.

Throughout the text we largely use scientific terminology that is well established in basic comparative historical linguistics, but for the sake of those readers who are not well versed in this subject it makes sense to provide a few concise definitions for some of the most important terms. Two or more languages are considered to be genealogically related if they may be traced, through historic sources or comparative reconstruction, to a single (attested or hypothetical) ancestral language. Such common ancestry is demonstrated through sets of cognates – full words or single morphemes (roots, affixes, etc.) inherited by the compared languages from their ancestor. As a rule, cognacy of specific items is proven by establishing regular (recurrent) sound correspondences which are common across multiple sets of cognates rather than isolated ones, thus ruling out the possibility of accidental resemblance (e.g. English t – German z, as in two – zwei, ten – zehn, to – zu, etc.). Since a genealogic relationship of languages is relative rather than absolute (i.e. some languages may be more closely related to each other than others, just like biological species), there exist various methods to measure degrees of relationship; most of them are lexicostatistical in nature, using fixed lists of the so-called basic vocabulary (words that represent some of the most common and universal concepts, such as body parts, simple natural objects, essential vital functions, etc.) to determine the relative proximity between languages by the percentages of cognates on these lists. For living languages, basic vocabulary wordlists may be collected from text corpora or dictionaries, or directly from native speakers; for proto-languages, they have to be reconstructed by applying a special onomasiological methodology (choosing the optimal candidate for the corresponding meaning on the wordlist, e.g. *twa- for Germanic ‘two’ on the basis of English two and German zwei, etc.).