A Bayesian phylogenetic study of the Dravidian language family

Abstract

The Dravidian language family consists of about 80 varieties (Hammarström H. 2016 Glottolog 2.7) spoken by 220 million people across southern and central India and surrounding countries (Steever SB. 1998 In The Dravidian languages (ed. SB Steever), pp. 1-39: 1). Neither the geographical origin of the Dravidian language homeland nor its exact dispersal through time are known. The history of these languages is crucial for understanding prehistory in Eurasia, because despite their current restricted range, these languages played a significant role in influencing other language groups including Indo-Aryan (Indo-European) and Munda (Austroasiatic) speakers. Here, we report the results of a Bayesian phylogenetic analysis of cognate-coded lexical data, elicited first hand from native speakers, to investigate the subgrouping of the Dravidian language family, and provide dates for the major points of diversification. Our results indicate that the Dravidian language family is approximately 4500 years old, a finding that corresponds well with earlier linguistic and archaeological studies. The main branches of the Dravidian language family (North, Central, South I, South II) are recovered, although the placement of languages within these main branches diverges from previous classifications. We find considerable uncertainty with regard to the relationships between the main branches.

Keywords: BEAST 2; Bayesian phylogenetic inference; Dravidian; dating; language phylogeny.

Figure 1.

Map of the Dravidian languages in India, Pakistan, Afghanistan and Nepal adapted from Ethnologue [2]. Each polygon represents a language variety (language or dialect). Colours correspond to subgroups (see text). The three large South I languages, Kannada, Tamil and Malayalam are light red, while the smaller South I languages are bright red. Languages present in the dataset used in this paper are indicated by name, with languages with long (950 + years) literatures in bold.

Figure 2.

A NeighborNet visualization of lexical differences. The NeighborNet network identifies three groups, going in clockwise direction starting from Telugu in the upper left corner: South II + Central (Telugu, Koya, Kolami, Gondi, Parji, Ollari Gadba and Kuwi), North (Kurukh, Malto and Brahui) and South I (Tulu, Tamil, Malayalam, Yeruva, Kodava, Kannada, Badga, Toda, Kota and Betta Kurumba). Colour-coding gives subgroup affiliation: red, South I; blue, Central; purple, North; yellow, South II.

Figure 3.

MCC tree summary of the posterior probability distribution of the tree sample generated by the analysis with the relaxed covarion model with relative mutation rates estimated. Node bars give the 95% highest posterior density (HPD) limits of the node heights. Numbers over branches give the posterior probability of the node to the right (range 0–1). Colour coding of the branches gives subgroup affiliation: red, South I; blue, Central; purple, North; yellow, South II.

Figure 4.

Analysis without monophyletic constraints on the North, South I and South II subgroup. MCC tree summary of the posterior probability distribution of the tree sample, relaxed covarion model with relative mutation rates estimated. Node bars give the 95% highest posterior density (HPD) limits of the node heights. Numbers over branches give the posterior probability of the node to the right (range 0–1). Colour coding of the branches gives subgroup affiliation: red, South I; blue, Central; purple, North; yellow, South II.

Figure 5.

Analysis with additional monophyletic constraints on South I + South II. MCC tree summary of the posterior probability distribution of the tree sample, relaxed covarion model with relative mutation rates estimated. Node bars give the 95% HPD limits of the node heights. Numbers over branches give the posterior probability of the node to the right (range 0–1). Colour coding of the branches gives subgroup affiliation: red, South I; blue, Central; purple, North; yellow, South II.

Figure 6.

Frequency plots of the age of the Dravidian language family (tree height) for all models, with age in years from the present. Analyses are ordered by marginal log-likelihood, table 3, and electronic supplementary material, table S2. Models of cognate gain and loss included are (i) CTMC; (ii) CTMC plus γ rate heterogeneity with four rate categories (CTMC + γ); (iii) binary covarion (covarion); (iv) stochastic Dollo (Dollo). Models of tree rate variation are (i) strict clock; (ii) relaxed clock with rates sampled from lognormal distribution. Models of concept rate variation are (i) relative mutation rate estimated for individual concepts (word rates); (ii) no relative mutation rate estimated (overall rate).