A striking recent paper by Quentin Atkinson ("Phonemic Diversity Supports a Serial Founder Effect Model of Language Expansion from Africa", Science 4/15/2011) has been the subject of a lot of discussion recently. Its abstract:
Human genetic and phenotypic diversity declines with distance from Africa, as predicted by a serial founder effect in which successive population bottlenecks during range expansion progressively reduce diversity, underpinning support for an African origin of modern humans. Recent work suggests that a similar founder effect may operate on human culture and language. Here I show that the number of phonemes used in a global sample of 504 languages is also clinal and fits a serial founder–effect model of expansion from an inferred origin in Africa. This result, which is not explained by more recent demographic history, local language diversity, or statistical non-independence within language families, points to parallel mechanisms shaping genetic and linguistic diversity and supports an African origin of modern human languages.
This paper's premises are intriguing, if far from obvious, and its results are pretty compelling:
However, I have some concerns about what lies in between the assumptions and the results, especially concerning the way "Total Phoneme Diversity" is estimated. This measure gives (what seems to me to be) excessive weight to certain features, ignores syllable structure, and (as a result) is heavily influenced by a few areal characteristics that are as likely to be innovations as survivals.
Let's start by looking at what Prof. Atkinson did. From the Supporting Online Material:
Data on phoneme inventory size were taken from the World Atlas of Language Structures (WALS – available online at http://www.wals.info/) (S1-S4) together with information on each language’s taxonomic affiliation (family, subfamily and genus) and geographic location (longitude and latitude). WALS contains information on three elements of phonemic diversity – vowel (S2), consonant (S3) and tone (S4) diversity – in a total of 567 languages. Due to uncertainty in ascertaining exact inventory counts across languages, the WALS data are binned into ranges for vowel (small [2-4], medium [5-6], large [7-14]), consonant (small [6-14], moderately small [15-18], average [19-25], moderately large [26-33], large [34+]) and tone (no tone, simple tone and complex tone) diversity. Uncertainty associated with these diversity assignments is only expected to weaken any clinal relationship with geography. WALS values for the three items were standardized (subtracting the mean value and then dividing the difference by the standard deviation) so that they were on comparable scales (mean = 0, standard deviation = 1). The standardized scores were then averaged to produce a measure of total phonemic diversity in each language.
Here is a boxplot giving the distributions of scores by region:
Given the continent-level distributions, it's not surprising that a plot of individual languages by distance from a putative origin in Africa falls off in a convincing way:
As the plot's caption explains,
Distance from the origin alone explains 30% of the variation in phonemic diversity (fitted line; r = –0.545, n = 504 languages, P < 0.001) and 19.2% of the variation after controlling for modern speaker population size.
The text tells us that
The relationship also holds for vowel (r = –0.394, P < 0.001), consonant (r = –0.260, P < 0.001), and tone diversity (r = –0.391, P < 0.001) separately.
But there's something about Atkinson's "Total Phoneme Diversity" that should strike you as odd. Tone, vowel, and consonant "diversity" are weighted equally, although the numbers of alternatives and the contribution to syllable- or word-level "diversity" are radically different in the three cases. Thus losing a single tone would generally reduce "Total Phoneme Diversity" by as much as losing about 10 consonants would. Worse, the lost tonal feature would probably give us only one pair of phonological "alleles" per syllable, at most, while the consonant choices would probably be available (to different extents in different languages) in several places in a syllable, so that the equivalent reduction in "phonemic variants" — seen as alternative symbols in different functional locations — might be numbered in the hundreds.
So let's take a closer look at how "Total Phoneme Diversity" was calculated. The data comes from M. Haspelmath, M. S. Dryer, D. Gil, B. Comrie, Eds., The World Atlas of Language Structures Online (Max Planck Digital Library, Munich, 2008), where (for plausible typological reasons) the phonological inventories of up to 567 languages are treated in a coarsely granular way.
The WALS tone inventories are divided into three classes, "no tones", "simple tone systems", "complex tone systems". After Atkinson's "normalization" (subtraction of the mean and division by the standard deviation) the three possible values for tone-inventory size turn into the numerical values -0.769, 0.547, and 1.862. Given that the "Total Normalized Phoneme Diversity" is the average of tone, vowel, and consonant measures, the contribution of the three possible values of "Normalized Tone Diversity" to the total will be -0.256, 0.182, 0.621.
|No tones||(307 languages)||diversity = -0.769|
|Simple tone system||(132 languages)||diversity = 0.547|
|Complex tone system||(88 languages)||diversity = 1.862|
Here's corresponding map:
Here are the consonant inventories:
|Small||6-14||(91 languages)||diversity = -1.554|
|Moderately small||15-18||(121 languages)||diversity = -0.717|
|Average||19-25||(182 languages)||diversity = 0.120|
|Moderately large||26-33||(116 languages)||diversity = 0.958|
|Large||33+||(53 languages)||diversity = 1.795|
And the corresponding consonant map:
The WALS vowel inventories:
|Small (2-4)||(93 languages)||diversity = -1.235|
|Average (5-6)||(288 languages)||diversity = -0.485|
|Large (7-14)||(183 languages)||diversity = 1.390|
And the corresponding vowel map:
It was plausible for the people who put the WALS data together to bin the phonological inventories so coarsely. (Ian Maddieson wrote the relevant chapters, but I suppose that the decision about how to quantize inventories was a joint one.) For a start, this coarse quantization avoids a lot of detailed argumentation about exactly how to analyze specific languages, like our recent discussion over whether /hw/ is a doubly-articulated consonant (and thus an addition of one to the consonant count) or a sequence of two consonants (and thus omitted from the count). I'm not sure that I would have made the same choice, but anyhow, it's done, and the facts on the WALS ground thus limited Atkinson's options.
However, this combination of coarse binning into ranges, for functionally-defined subsets of elements with radically different numbers of members, seems to me to be much more problematic for Atkinson's purposes. It's as if a human genomic survey made geographically localized counts of the number of alleles involved in color vision and in blood physiology, divided each set of counts into a few bins ("a little variation", "a medium amount of variation", "a lot of variation"), standardized the binned counts for each functional class separately, and averaged the results, thus giving as much weight to each color-vision variant as to several orders of magnitude more blood-physiology variants. This might be OK, but choosing to give this kind of boost to features that happen to be enriched in one region or another will obviously push the results around by a considerable amount.
And indeed in this case, a few areal features (which might well be innovations rather than retained characteristics) have an outsized effect on the results. For example, nearly all the languages of sub-Saharan Africa have lexical tone. To quantify the effect of that single feature, I downloaded the WALS data and crudely segregated the data for Africa, Europe, and South America (using latitude and longitude rectangles, since continent is not noted in the database). On that basis, the average "Normalized Tone Diversity" of those three regions was 0.934 for Africa, -0.450 for South America, and -0.637 for Europe — a difference of more than one and a half standard deviations, reflecting an areal distribution of what is basically a single phonological feature.
Dividing by three and subtracting, this would shift Africa by -0.524 in "Total Phonemic Diversity" relative to Europe, and by -0.461 relative to South America, which is roughly what the overall mean regional differences are.
There are some similar effects for vowels — many African languages use vowel nasality and/or the "advanced tongue root" feature to double or quadruple their vowel inventories, while keeping their syllable structure simple. Such things happen to vowels elsewhere in the world, but (in most places) not as often; roughly in compensation, languages elsewhere tend on average to have more complex syllable structures.
Overall, as I noted above, it's not at all obvious what we ought to count in creating data for such a project. Should we generally count long and short vowels as separate entities or as sequences? What about monophthongs and diphthongs? Or nasal and oral vowels? If we delete syllable-final nasals (retaining syllable-initial nasals) and instead encode the same information in nasality of an inventory of 4 vowels, it looks like we've doubled the number of vowels — moving us from a small (2-4) number of vowels to a large (7-14) number of vowels — while leaving the consonant inventory unchanged.
These questions (and many others having to do with syllable and morpheme structure) don't just add noise to the data. Many such features are "areal" — spread over geographical areas of related (and even unrelated) languages, and also spread over local descriptive practices among linguists. The areas in question often approach continental size, and thus these phenomena (and our choices about how to code them) can have a big influence on the outcome of algorithms that look for world-wide clines in the "diversity" measures that result from these choices.
Finally, it's worth adding a few words on a key assumption of this work, laid out with admirable clarity in the opening sentences:
The number of phonemes—perceptually distinct units of sound that differentiate words—in a language is positively correlated with the size of its speaker population (1) in such a way that small populations have fewer phonemes. Languages continually gain and lose phonemes because of stochastic processes (2, 3). If phoneme distinctions are more likely to be lost in small founder populations, then a succession of founder events during range expansion should progressively reduce phonemic diversity with increasing distance from the point of origin, paralleling the serial founder effect observed in population genetics (4–9).
Reference (1) is Jennifer Hay and Laurie Bauer, "Phoneme inventory size and population size", Language 2007. The data reviewed in that paper is certainly consistent with the hypothesized "founder effect", as exemplified by their Figure 1:
I was initially somewhat skeptical of this result, but the effects seem to be quite robust over data sets, modes of analysis, correction for various possible confounding factors, etc. The most plausible remaining possible artefact seems to be the issue that Hay and Bauer describe this way:
A reviewer of this paper suggested that it could be possible that the number of ‘phonemes’ in a language tends to increase as the language is studied, and that languages spoken by more speakers tend to receive more attention. This is an intriguing suggestion, which would provide a sociological explanation for our observed effect. However this portion of our analysis has reduced each language family to one data point – investigating the mean population size and phoneme inventory by language family. This eliminates the possibility that a few highly studied language families (such as Indo-European) might be driving the effect in this way. The fact that the correlation is robust both across languages and language families suggests strongly that there is something here that requires explanation.
As is often the case, this territory has previously been explored by the eminent American linguist, Dave Barry (Dentists in Paradise), including a discussion of the concomitant increase in average morpheme length:
The Hawaiian language is quite unusual because when the original Polynesians came in their canoes, most of their consonants were washed overboard in a storm, and they arrived here with almost nothing but vowels. All the streets have names like Kal'ia'iou'amaa'aaa'eiou, and many street signs spontaneously generate new syllables during the night.
Let me close with a plug for the forthcoming report, "Fashion Diversity Supports a Serial Founder Effect Model of Expansion From North-Central France", extending my own seminal work in "The Hunt for the Hat Gene". I can't reveal too much, since the work is still under embargo, but it has been conclusively established that smaller populations have fewer styles of dress (including fewer kinds of hats), so that we should expect the global distribution of diversity in clothing and headgear to show the effect of population bottlenecks in the migration of humans from our ancestral home in the valley of the Seine.