When I had lunch with two phoneticians the other day, the question came up whether there is a correlation between the size of the phoneme inventory of a language and its population size. Neither of us knew, and we started to speculate in which direction it should go. My guess was that larger languages should have smaller inventories and vice versa, while the two expert expected the opposite. One of the reasons for my conjecture was the observation by Lupyan and Dale (2010) in Plos ONE (Lupyan G, Dale R (2010) Language Structure Is Partly Determined by Social Structure. PLoS ONE 5(1): e8559. doi:10.1371/journal.pone.0008559) that larger languages tend to have less morphological complexity (and vice versa). Lupyan and Dale speculate that this may be due to the fact that large language communities comprise a larger proportion of L2 learners. As L2 learners – I am simplifying here, of course – have a hard time acquiring all morphological complexities, they cause an overall simplification of their adopted language in the long run.
Well, as most of us know from own frustrating experience, mastering the sound inventory of a foreign language as an adult is even harder than mastering the morphology. By the same logic, one should assume that larger languages, having more L2 speakers, have a smaller sound inventory.
The following day I learned that this issue has been resolved already back in 2007, in a Language article by Jennifer Hay and Laurie Bauer (Phoneme inventory size and population size, Language 83(2), 388-400). They found exactly the opposite of what I expected – so the phoneticians were right and I was wrong.
As I had started to explore this issue on my own already, I want to share my findings nevertheless, even though its no news anymore.
I went to the data base of the Automated Similarity Judgment Program (http://wwwstaff.eva.mpg.de/~wichmann/ASJPHomePage.htm), which consists of 40-item Swadesh lists from more then 5,000 languages and dialects. All words are given in a uniform, rather coarse grained phonetic transcription. Conveniently, the data base also reports population size according to Ethnologue. As a crude measure of the sound inventory of a language, I simply counted the number of different symbols occurring within its Swadesh list. The method is quick and dirty because (a) the transcriptions are phonetic rather than phonemic, so I count different sounds rather than different phonemes and (b) there is no guarantee that all sounds of a language occur within such a short word list, but it should give us a rough idea at least.
Plotting population size against sound inventory size looks like this:
The tendency is obviously the same that Hay and Bauer report: small languages tend to have small sound inventories and vice versa. Here is some statistics if you care for numbers rather than images:
- correlation between log population size and sound inventory size: 0.26, 95% confidence interval: (0.24,0.28), p< 2e-16.
- a linear regression of sound inventory size against log population size gives R-squared = 0.067. When controlling for the size of the word list (because ASJP has many missing entries, and sometimes we have more than one word per concept), R-squared is still 0.064.
So, to put it in a crude way: small languages tend to have few sounds and many morphemes, and large languages behave the opposite way. Now the question arises: are phonology and morphology independently affected by population size, or is there something to this popular myth after all that there is a tradeoff in complexity between different levels of linguistic structure?
P.S. It turns out that this is not just old news, but very old news. Basically the very same analysis, along with a lot of other interesting stuff, can be found in a 2011 paper by Wichmann, Rama and Holman: http://www.degruyter.com/view/j/lity.2011.15.issue-2/lity.2011.013/lity.2011.013.xml
P.P.S. Wichman et al. did their study with version 13 of ASJP, while I used version 15. So my work was not entirely in vain. Also, they show that there is a negative correlation between population size and average word length. Since there is also a negative correlation between average word length and sound inventory size, the causal chain might be something like
large population -> shorter words -> more phonemes
However, when I control for average word length (in the ASJP word list), the correlation between log population size and sound inventory size is still signficant, r=0.18, R-squared=0.03. With the caveat that (unweighted) average word length in a 40-item Swadesh list is a very crude estimate of average word length in actual language use, it seems that the effect of population size on sound inventory size cannot fully be mediated by word length.
It seems the question may be ill-formed. Why should population size have any effect on phonological inventory? I have no research I can cite, but experience tells me that L2s have a minimal and marginal effect in shaping the languages they learn. E.g. a handful of Chinese phrases like “long time no see” have made it into English; however, native speakers around the world have not incorporated any Chinese L2 pronunciation into English. Locally one might see this effect. However, sound substitutions for “difficult” sounds may or may not reduce the size of the inventory. E.g. Chinese struggle with [z] and replace it with [ʐ]. In such cases inventory has not reduced, merely shifted. But again, that’s likely moot for sociolinguistic reasons. “Inferior” varieties defer to “superior” and 2nd generation L1s conform to native standards.
Given that the correlation holds across language families and not within them, is our sample really big enough for results to be significant? That is, wouldn’t it be the number of families being analyzed that is the most relevant way of measuring sample size in this case, rather than the number of languages? If so, it’s worth bearing in mind that the number of language families in the world is relatively small.
David Erschler hypothesized above that Austronesia and Papuan languages might themselves throw off the statistics – but even if this is not the case, might there not be other random, coincidental distributions among the world’s relatively small number of language families?
You are probably right. When I compute the average log population and average sound inventory size per family, the correlation goes down to 11.7%, and the p-value is 9.4%. So it is not significant anymore.
I see. Still, it’s an interesting avenue of research. I’ll be keeping an eye out to see if the research community can uncover or further clarify anything down the line.
This debate continues in a more recent paper, but with a quite sobering conclusion:
Moran, S., McCloy, D., & Wright, R. (2012). Revisiting the population vs phoneme-inventory correlation. Language 88(4), 877–893.
The authors argue that the Hay/Bauer results are a statistical artifact and that, once genealogical relationships between languages are filtered out as a contributing factor, there is actually litte or no correlation between phoneme inventory and population size.
Thanks for the reference! After a first brief glance at Moran et al., it seems that my previous post is in accordance with their finding – we do not find the alleged correlation within genetic units. I can’t even think of a reasonably simple mechanism that would explain the inter-family correlation that I found, so the whole thing might in fact be spurious.
Here is an interesting twist to the story. The effect is robust *across* language families. So if you pick two language families at random, and pick on language from each family, the odds are that the language with the larger population will have a larger segment inventory. However, it does *not* hold within families. The within-family correlation is not significant in most cases, and where it is significant, it is positive and negative in equal proportion. The average within-family correlation (averaged over all families) is almost exactly 0. So if you pick a family at random and then two languages from this family, population size is not a predictor for segment inventory size.
Differences in population size are historically quite young, certainly younger than most language families. If there were a causal link between population size and segment inventory size, we would expect to find a correlation within families but not across families. What we do find is exactly the opposite. So the whole thing remains a puzzle.
At the end of a forthcoming paper I have speculated a little about possible diachronic reasons for such correlations, with a very brief discussion of the papers mentioned here.
The paper can be found here
Thanks for the link!
But won’t the ASJP simplified transcription considerably distort the picture?
Sure, there are a lot of phonetic contrasts that the ASJP transcription does not capture (tones, oral/nasal vowels, vowel length is not transcribed uniformly, distinction between different r-sounds, pharyngal vs. uvular fricatives …). Still, I don’t think the correlation with population size is an artifact of the transcription. Otherwise we would have to assume that small languages have a lot of distinction within ASJP classes and large languages between ASJP classes. And this would be distinctly odd…
Well, one could may be argue that the definition of sound classes in ASJP is somewhat biased towards better known languages, but this is not very plausible indeed.
Another possible contribution to this effect might be that very numerous Austronesian and Papuan languages tend to have both relatively small speaker numbers and small phonemic inventories.
Leaving out all Austronesian and Papuan languages reduces the correlation slightly (to 0.22), but its still highly significant.
I haven’t seen the articles by Wichman et al. mentioned in the p(p)s. but another interesting “old” perspective on this, it seems to me, is the research by David Nettle on the relationship between phoneme inventory size and what he calls “communicative efficiency” (see his “Segmental inventory size, word length, and communicative efficiency”, Linguistics 3 (2), pp. 359-367). Nettle shows that there is a tradeoff between word length and inventory size, such that the larger the phoneme inventory is, the smaller the average length of a word in a given language. This
means, a fortiori, that there is a functional relationship between articulatory ease and perceptual salience, which leads to an average word length of 7 +/- 2 segments per word in the sample of languages he studied, thus corresponding well to average capacities of working memory (the “magical number 7 +/- 2”!). It also chimes well with a famous dictum by Roman Jacobson (1929), who held that “the more diffuse the geographical range of a language, the simpler the system had to be, as ease of learning and discrimination under accent diversity would be preeminently important”.
If the correlation Bauer et al. propose actually holds, it would mean that if you have a large population size, you compensate the encountered communicative variability by increasing the phoneme inventory, which in turn decreases the length of a word, so that you end up with a cluster/dipthong/tone-rich monosyllabic system. (Unless, of course, it is your segmental inventoy which influences population size ;-)) But then, other work by Nettle (“Linguistic fragmentation and the wealth of nations: the
Fishman-Pool Hypothesis Reexamines”, Economic Development and Cultural Change 2000: 335-348) has stressed that there is no clearcut correlation between linguistic homogeneity and economic success in a given area.