Typological Linguistics and Population Genetics: A Synthesis or a Controversy
Trends in Cognitive Sciences Vol 16, Issue 3, March 2012, Pages 167–173 http://dx.doi.org/10.1016/j.tics.2012.01.007
Tools from Evolutionary Biology Shed New Light on the Diversification of Languages
Stephen C. Levinson, and Russell D. Gray
Computational methods have revolutionized evolutionary biology. In this paper we explore the impact these methods are now having on our understanding of the forces that both affect the diversification of human languages and shape human cognition. We show how these methods can illuminate problems ranging from the nature of constraints on linguistic variation to the role that social processes play in determining the rate of linguistic change. Throughout the paper we argue that the cognitive sciences should move away from an idealized model of human cognition, to a more biologically realistic model where variation is central.
As I have emphasized many times on this blog, only an interdisciplinary approach is capable of generating an optimal theory of modern human origins. We are moving away from the era when archaeology and paleobiology dictated the course of thinking about modern human origins. Global analyses of genetic variation have put a firm stake in the ground as a premier source of testable hypotheses for ancient population history. But they, too, failed to provide an accurate picture of modern human evolution, as geneticists’ brainchild, out-of-Africa-with-Complete-Replacement, was falsified by the growing evidence of “admixture” from archaic hominins. Linguistics is getting pulled into an overall troubled synthesis, as “hand-made” genealogical classifications of the world’s 6000+ languages are being complemented by quantitative analyses of the distribution of typological features. Stephen Levinson and Russell Gray provide a short overview of the current state of affairs at the intersection of typological linguistics and population genetics. In it, they celebrate the successes of methods derived from biology in driving better understanding of linguistic variation, while highlighting various problems and opportunities, too.
I’m generally somewhat skeptical about the “success stories.” I blogged about the unsatisfactory outcomes of Bayesian phylogenetics when it was applied to a) the dispersal of the Indo-European language family and b) the reconstruction of the evolution of Austronesian sibling terminologies. (As a reader informs me, the former has recently caused an outrage in some circles sympathetic to traditional Indo-European linguistics and the broadly accepted consensus between archaeologists and linguists around the Kurgan (Pontic Steppe) model of Indo-European dispersals.) And I consider a fiasco Quentin Atkinson’s attempt to model the decay of phonetic inventory sizes on the gradient of diminishing genetic diversity with increasing distance from Africa. However, these “flops” are not void of significance because they allow one to reverse engineer a critique of the underlying biological and genetic models. If these methods do not work for languages, they are unlikely to work for genes. And, indeed, recent empirical evidence (“admixture” of non-Africans with Neandertals and Denisovans) falsified the Out-of-Africa-with-Complete-Replacement-Model, which at one point was assigned a very high posterior probability score (see also here).
I’m also a firm believer in quantitative methodologies and am confident that there are and/or will be occasions when they are applied to languages with breakthrough results.
In this context, one case study highlighted by Levinson & Gray (2012) deserves a special mention. It is now customary to think (see also here) that Joseph Greenberg’s linguistic classifications (Amerind, Eurasiatic, Indo-Pacific, Khoisan, Niger-Congo, etc.) are flawed (to a various degree of severity) because they obscure true genealogical relationships between underlying families by lumping them into higher-order phylogenetic units on the basis of spurious lexical and morphological similarities. At the same time, Greenberg’s contributions to linguistic typology are hailed as a major breakthrough in synchronic linguistics. But Greenberg’s diachronic and synchronic studies were always two sides of the same coin. He called his comparative method “taxonomic” as it relied not on a pairwise comparison between languages but on scoring similarities in sound and meaning over a wide array of languages mostly found across continuous geographic areas. He took as a default assumption the fact that these formal similarities are driven by common descent, not convergence or contact. This assumption made a significant number of phonetic units in world languages (e.g., the m/n pronouns in American Indian languages) look immune to the vagaries of history and, hence, approaching the status of universals. Naturally, his follower, Merritt Ruhlen, began identifying phonetic strings that have not changed since the proto-language of Homo sapiens. A pair of French amateur scholars, Pierre Bancel and Alain Matthey de l’Etang, sieved through thousands of reduplicated kin terms in world languages arguing that the pervasive similarities between kaka, mama, tata and other words derive not from multiple independent origins in baby talk but from the fact that they are retained, virtually unchanged, from the proto-language of Homo sapiens.
One of Greenberg’s breakthrough typological findings is that the position of Verb vis-a-vis its Object controls the position of such secondary parts of the sentence as adpositions, adjectives, determiners, etc. Advanced on the basis of just 30 languages, Greenberg’s universals of word order received support from Matthew Dryer’s study (“The Evidence for Word Order Correlations,” Language Typology 15 (2011), 335-380) of word order tendencies in a large sample of 1000 languages. Levinson & Gray (2012), however, report that Bayesian analysis fails to support the existence of systematic functional linkages between features. The word order features show a pattern of association specific to the language family. For example, only Austronesian showed strongly coupled changes between numeral-noun and genitive-noun orders. If true, this refutation of one of Greenberg’s most famous linguistic universals indicates that his studies in typology may be just as flawed as his linguistic classifications and for exactly the same reason, namely Greenberg’s belief in the existence of innate (cognitive) constraints on linguistic variation. In this regard, Greenberg approached Chomsky and no wonder Chomsky endorsed Greenebrg’s typological findings as “descriptive generalizations that should be derived from principles of UG [Universal Grammar – G.D.]” (Chomsky, Noam. 1998. Minimalist Inquiries: The Framework. MIT Occasional Papers in Linguistics 15.) Levinson & Gray’s critique, then, falls under the same category as Everett’s critique of Universal Grammar from the point of view of Piraha language facts. A sophisticated quantitative methodology arrives at the same conclusion as a deeply immersive linguistic ethnography.
Another interesting part of Levinson & Gray (2012) is a summary of arguments advanced against Quentin Atkinson’s claim that phonemic diversity decays with increasing distance from Africa and thus mirrors the pattern observed in worldwide genetic diversity.
“In general, the areas of the globe that were most recently colonized show fewer phonemes in the local languages, whereas the areas that have hosted modern humans for millennia (particularly sub-Saharan Africa) still, on average, use the most phonemes. There are some obvious exceptions to the pattern Atkinson reports. Some of the world’s largest phoneme inventories are found in the Caucasus and among Na Dene languages in North America, which was colonized relatively late. It is also not clear whether the serial founder effect model applies to phoneme inventories. Human populations are normally genetically diverse but speech communities typically have little variation in their phoneme inventories. In cases where there have been well attested serial founder events, such as the Austronesian expansion into the Pacific from Taiwan, the largest phoneme inventories are not located in Taiwan but rather in the languages of New Caledonia and the Loyalties . Although the general pattern Atkinson found has been replicated using other data and methods [29–31], the findings are vulnerable to slight changes in assumptions , other gradients can be found and a distance-from-Africa gradient might well be due to other causes .”
The first sentence demonstrates that Levinson & Gray (2012) are content with reiterating one of Atkinson’s key premises, namely that we know that places like South America were colonized recently, while Sub-Saharan Africa has hosted modern humans for millennia. This is the same assumption that a decade earlier Daniel Nettle (“Linguistic Diversity of the Americas Can Be Reconciled with a Recent Colonization,” Proc Nat Acad Sci 96 (1999), 3325-3329) turned into an argument against Johanna Nichols’s finding (“Linguistic Diversity and the First Settlement of the New World,” Language 66 (3): 475-52) that phenomenal linguistic diversity in the Americas as measured by the number of independent stocks contradicts the archaeological consensus advocating for a recent (12,000 YBP) colonization of the New World. Nettle (and earlier Dixon) linked the perception of Africa and Europe being the areas occupied by modern human the longest with the objective reality that linguistic diversity in these areas is the lowest to produce a kind of law of inverse correlation between linguistic diversity and the length of occupation: when a population colonizes a new habitat, its linguistic diversity first rises steeply but then declines steadily. (The logic is similar to the way geneticists interpret patterns of linkage disequilibrium: Sub-Saharan Africans has low LD and American Indians have high LD because Africans had enough time to break down the chunks of homozygosity after the founding bottleneck, whereas American Indian populations still bear the “scar” of the population depression caused by the founding Beringian bottleneck.) Africa and Europe have low stock diversity because forces of homogenization (e.g., absorption of forager substrates by agriculturalists such as Bantu and Indo-Europeans) have had enough time to bring stock diversity down, America has high stock diversity because, by 1492, it did not have enough time to eliminate it through language contact and extinction.
The irony is that, while Nettle argued for inverse correlation between linguistic stock diversity and population age, Atkinson argued for direct correlation between phonemic diversity and population age. Atkinson could have used Nettle’s logic and argued that phonemic diversity in the Americas is the lowest, hence it’s the continent with the longest history of modern human occupation because, as time goes by, linguistic diversity of whatever kind tends to decline. But he did not reference Nettle’s argument because it would have made the gradient of phonemic diversity compatible with an out-of-America scenario. He clearly could not allow that because his goal was to show the signal of an out-of-Africa expansion detected by geneticists manifests itself in the same way in a pattern of linguistic diversity. Nettle, on the other hand, could not agree with Nichols because Nichols’s built a case for questioning the picture archaeology painted for the colonization of the New World. This demonstrates that it is the assumption of which continent was settled first and which one last that drives Nettle’s and Atkinson’s interpretations of linguistic evidence, not the actual data. The actual data is fitted into the intellectual mold. And Levinson & Gray (2012) reinforce this regrettable practice.
Unlike Nettle’s debate with Nichols, Atkinson’s model received lots of peer attention. Scholars have raised quite a few objections, and Levinson & Gray (2012) enumerated some of them.
1. There are empirical cases when languages with largest phonemic inventory sizes are not basal in the language family phylogeny. E.g., in the Austronesian family the largest phoneme inventories are not located in Taiwan but rather in the languages of New Caledonia and the Loyalties.
2. Large phonemic inventories pop up in unexpected places (e.g., Na-Dene in North America and the languages of the Caucasus), and, in the case of Na-Dene, the consensus is that it is a newly derived family.
3. If the phonemic diversity gradient mimics the colonization route, then America should be able to provide a good case study, as it is firmly believed that it was colonized from Beringia and this migration happened much more recently than the founding migration out of Africa. We therefore should see a better signal of the dependency postulated by Atkinson in the New World than anywhere else. But in reality in the New World phonemic diversity decreases from west to east, not from north to south.
“There is a highly significant correlation between these calculated distances and the phonological index (R2 = .209, p < .0001) reflecting lower complexity at more easterly locations. In a stepwise multiple regression with distance from the Bering Strait, latitude, and “corrected” longitudinal distances as predictors of the summed consonant/syllable index, the approximated east-west distance is the first to enter, a significant improvement to the model is made by adding latitude (cumulative R2 = .336, p < .00001), whereas distance from the Bering Strait makes no contribution (p = .5239)” (Maddieson et al. 2011. Geographical Distribution of Phonological Complexity,” Linguistic Typology 15, 276).
3. The outcome of Atkinson’s analysis was pre-selected by the very inclusion of the distance criterion into the model. As Jaeger et al. (2011. “Mixed Effect Models for Genetic and Areal Dependencies in Linguistic Typology,” Linguistic Typology 15, 311) argued,
“[T]he distance effect is assessed in the model for which adding distance to the model improved the model’s deviance the most compared to all other possible origins. This procedure is obviously biased to find an effect for distance to the origin: if there is such an effect for any hypothetical origin, it will find it.”
4. The notion of “phonemic diversity” is misleading and is not in any way parallel to the notion of “genetic diversity.” As Maddieson et al. (2011, 272) pointed out,
“A subgroup of speakers of a given language does not use a subset of the phonemes of the language, but all of them. Hence the basis for expecting an effect at the phonological level that is directly analogous to a founder effect in genetic terms is questionable. Moreover, a language’s inventory of tones and phonemes is a property that emerges from the language’s lexicon, rather than being a property that exists independently. To lose a phoneme simply by the act of founding a new population would entail highly selective loss of the particular lexical items containing that phoneme.”
(As a side remark, Levinson & Gray (2012, 4), for some reason, list “founder effect” as a clear parallel between biological and linguistic evolution. In reality, it is far from being the case at least when it comes to “phonemic diversity.” It could become a “parallel” if geneticists shifted to a population model whereby every subpopulation contains all the diversity contained in a metapopulation.)
5. The closest approximation to the notion of “genetic diversity” in linguistics is the measurement of the summed stock, phonemic, morphological, etc. diversities in a continent, but this was not modeled by Atkinson. Why did Atkinson not model linguistic diversities in a way that would be most similar to the way geneticists model genetic diversities? Perhaps this is because, if measured properly, linguistic diversity would not support the out-of-Africa model. Maddieson et al. (201, 272-273) explain,
“If it is the case that Africa held the greatest diversity of languages at the origin, and subsequent spread of modern humans to the rest of the world led to progressive reduction in their diversity, it is surprising that standard methods of linguistic comparison converge on the idea that there are few language families in Africa today. Although this consensus has been more recently challenged (see, e.g., Mous 2003, Sands 2009), even the suggested higher estimates for the number of language families in Africa are far lower than the number of families that exist outside Africa, and are lower than the number of language families in the Americas by itself in particular.
If there is a signal of an African origin in the phonological complexity of languages, we might expect that a similar signal might also be detectable in lexical, morphological, or syntactic patterns. So far, this has certainly not been the case. For example, all of the three most rare orders of Subject, Object, and Verb (VOS, OVS, OSV) are predominently found far from Africa (Feature 81A in WALS; Dryer 2011). In the WALS chapter on this issue all examples of the VOS order are outside Africa (including Madagascar, counted as part of South- East Asia by Atkinson). None of these less common orders occur anywhere in mainland Eurasia. Most cases of the OVS order are found in South America, where all six possible orders occur. If linguistic complexity in general declines with distance from Africa, this distribution would be unexpected.”
If feature complexity is chosen as an indicator of place of origin, 16 randomly picked WALS features pointed to many different homelands located on all continents, with Africa scoring the lowest on Bayesian Information Criterion (BIC) in the case of only one WALS 65 feature “perfective/imperfective marking” (Cysouw et al. 2011. Comment on ‘Phonemic Diversity Supports a Serial Founder Effect Model of Language Expansion from Africa’, Science 335, no. 6069, uppl. Mat., 26).
6. Phonemic inventory sizes are positively correlated with speaker community sizes. Africa, Europe, South Asia and Southeast Asia tend to have significantly higher speaker community sizes than America and the Sahul. When corrected for speaker community size, North America emerges as having the largest phonemic inventories (Cysouw et al. 2011, Suppl. Mat, 19).
7. There are other explanations to the observed pattern of high phonemic inventory sizes in Africa vs. the rest of the world. The size of the phonemic inventory may depend on word length (Nettle, David. 1998. “Segmental Inventory Size, Word Length, and Communicative Efficiency”, Linguistics 33 (2): 359-367; Wichmann et al. 2011. “Phonological Diversity, Word Length, and Population Sizes Across Languages: The ASJP Evidence,” Linguistic Typology 15, 177-197). Climatic/ecological environment may favor certain kinds of consonants and vowels. A crude test of this hypothesis conducted by Maddieson et al. (2011) captured at least just as much of the world’s phonemic variability as Atkinson’s model.
Scholars rejected Atkinson’s argument for linguistic evidence for an out-of-Africa expansion but they made sure to refer to his work as groundbreaking. As Levinson & Gray (2012, 3) phrased it,
“Atkinson’s application of evolutionary thinking is likely to stimulate many more efforts to explain global patterns of linguistic diversity by drawing on the parallels between the processes of biological and linguistic diversification.”
But I would argue that at this point in our knowledge of worldwide variation in genetic and linguistic traits “parallels” and “evolutionary thinking” are just not enough. Dixon and Nettle already saw a parallel between the gradient of increasing linguistic diversity from Africa and Europe to Papua New Guinea and America, on the one hand, and punctuated equilibrium in biology with very little proof that this parallel explains the nature of linguistic variation in the world. “Parallels” can be interpreted in wildly different ways. Genetic and linguistic systems are complex and it is unclear which part of the genetic picture is supposed to parallel which linguistic picture and why.
Peers have set a very low bar for Atkinson. They wished he was right but his evidence, logic and methodology are simply not up to par with what a robust theory of modern human origins written from cultural and linguistic evidence should look like. Levinson & Gray (2012) provide a useful summary of recent developments in the growing field of studies linking linguistics and biology/genetics, but it is not as analytical or critical as one would hope. The authors choose to celebrate the “parallels” without focusing enough on the challenges.