Goddard and Frohlich Respond to Atkinson

« previous post | next post »

The following is a letter written by Ives Goddard and Bruno Frohlich of the Department of Anthropology at the Smithsonian Institution in response to the paper by Q. D. Atkinson claiming that the distribution of speech sounds in the world's languages demonstrates a single point of origin for human languages in Africa. Mark discussed this paper here. The letter was submitted to Science, which declined to publish it.


Atkinson (1) claims that 19% of the variation in the size of the phoneme inventories in a sample of 504 languages from the World Atlas of Language Structures (WALS) (2), controlled for population size, is accounted for by distance along assumed ancient human migration routes and argues that this clinal distribution “fits a serial founder-effect model of expansion from an inferred origin in Africa,” This explanation rests on the unrealistic assumption that in population bottlenecks phonemes are likely to be lost according to the “same scenario” that brings about a loss of diversity in genetic makeup and material culture. Children have always learned to talk from a small number of speakers, regardless of the total size of the speech community, and no mechanism of language change is known that would plausibly explain why speakers should lose phonemes from a language they already speak just because the speech community becomes smaller, as Atkinson, offering no evidence, suggests “can” happen (3). This reduced population will expectably have fewer alleles per locus and less cultural knowledge at the margins, but there are no data that suggest that the number of phonemes will not stay the same until they are affected by the usual processes of language change. The fundamental scientific principle of uniformitarianism excludes positing for unattested ancestral languages mechanisms of change that have not been documented in the observable world.

As WALS states, “languages with larger or smaller inventories …. display quite marked regional disparities in their distribution.” Such features shared by contiguous, unrelated languages may define linguistic areas (4), and postulating the diffusion of areal influences (which Atkinson did not control for) in addition to inheritance within language families uncontroversially explains the correlation of phonological features and geography, such as the clustering of the 91 languages in WALS with small consonant inventories (6 to 14 consonants) in Oceania (47 languages, 29 on New Guinea alone) and South America (19). Number of consonants is a moderately robust measure of an objectively determined feature (though WALS has errors and gaps here), but the ranking of vowel quality and tone inventories in WALS lumps together phonological subsystems of widely differing complexity, and the “phonemic diversity” index produced by Atkinson’s averaging of normalized rankings of the size ranges in all three inventories does not correspond to a recognized feature of human language.

  • Q.D.Atkinson, Science 332, 346 (2011).
  • I. Maddieson, in The World Atlas of Language Structures Online, M. Haspelmath, M. S. Dryer, D. Gil, B. Comrie, Eds. (Max Planck Digital Library, Munich, 2008 [revised edition posted April 2011]).
  • Q.D.Atkinson, Supporting Online Material (http://www.sciencemag.org/cgi/content/full/332/6027/346/DC1), sect. 2.1.
  • M.S. Dryer, Studies in Language 13, 257 (1989); S. Thomason, in D. Gilbers, J. Nerbonne, J. Schaeken, Languages in Contact (Rodopi, Amsterdam, 2000), pp. 311–327.
Share:



35 Comments »

  1. secretseasons said,

    May 20, 2011 @ 1:50 pm

    Interested but non-expert here:
    No individual loses genes in the founder effect, and likewise no individual is postulated to lose phonemes in this analogous situation. But if you round up 1000 speakers of English you will get some diversity in the way those speakers speak the language; accordingly if you took only a subset (say 100) of those, you would get less diversity. This is the "loss" of phonemes and it is completely analogous to the "loss" of genes.

    Or am I misunderstanding the whole thing?

    [(myl) The effect that you've stated does not exist -- the "diversity in the way that [1000] speakers speak the language" — or a million speakers, supposing that they're speaking the same language — is not represented in any way at all in a phonemic inventory. That's a key point about a "phoneme inventory" — it abstracts away from phonetic diversity, and focuses only on the differences that distinguish words.

    Thus the fact that some English speakers pronounce goat with the vowel [o] while others say [əʊ] (and others say others things entirely for the same set of words) adds nothing whatsoever to the count of distinct vowels in English.

    It's true that different dialects may have phonemic splits or mergers, such as the English speakers who have only one category for the CLOTH, THOUGHT and PALM lexical sets, vs. those who preserve two or three categories. And it's possible that the WALS inventories for "large" as opposed to "small" languages may tend to minimize such mergers, because in the case of larger and better-documented languages it's more likely that some un-merged speakers have been documented, and an overall description may tend to include the distinction with a footnote to the effect that some speakers have lost it. But if this is true, it's a minor effect compared to sort of "diversity" you're talking about, which is absolutely excluded from the WALS data.

    So basically, yes, you're misunderstanding the whole thing.]

  2. John Cowan said,

    May 20, 2011 @ 2:28 pm

    I am unsurprised that Science declined to publish the above: it might expose the now-evident fact that general science journals are incompetent to tell good work from bad in historical linguistics. (Not that they should feed bad: few are so equipped at all. "Keine unter allen den Wissenschaften ist stolzer, edler, streitsüchtiger als die Philologie, oder gegen Fehler unbarmherziger.")

    Three cheers for Goddard and Frohlich for denouncing nonsense, and three cheers for you for publishing it.

  3. secretseasons said,

    May 20, 2011 @ 2:57 pm

    Thank you for the follow-up. Would you mind taking it a little further? How, exactly, are different dialects (let's say of English) represented in a phoneme inventory? A million English-speakers collected from all around Britain will use different phonemes, right? So is every different dialect represented separately?

    [(myl) The short answer to the question "How, exactly, are different dialects (let's say of English) represented in a phoneme inventory?" is "Not at all, except insofar as they have different numbers of systematic splits or mergers of categories".

    John Wells' description of the English language world-wide consists of 24 "lexical sets" specifying the vowel nuclei that all varieties of English. Thousands of papers have been written describing the many ways in which the pronunciation of these vowels vary across space, time, society, and linguistic context; and these only scratch the surface of the phonetic variation that's out there. But none of that variation changes the number of categories, unless it results in a systematic merger or split. (Another analyst might choose to omit some of Wells' categories, on the grounds that they play no role in some region's varieties, or to add some other contexts, for instance splitting off the case where the PRICE set has a following voiced vs. voiceless consonant -- PRICE vs. PRIZE -- because this does play a role in context-conditioned variation -- but that doesn't in itself change the number of phonemic categories either.)

    10 of the 24 "lexical sets" involve combinations of vowels with /r/ or sequences of vowels and glides, which are excluded by the definition of "vowel inventory" in WALS, leaving 14. My own variety of English, for what it's worth, merges two of the remaining sets, leaving 12 -- but that doesn't matter, because the rather crude WALS classification assigns English to the category of languages with a "large" vowel inventory, counted as "7-14".

    Note that WALS (where Atkinson got his data) only distinguish three sizes of inventories: small (2-4 in the case of vowels), average (5-6), and large (7-14).]

  4. John Lawler said,

    May 20, 2011 @ 3:22 pm

    @seacretseasons: Probably you should start by looking at what Ian Maddieson said about consonant inventories in WALS, since that's where the data come from.

    [(myl) You should also read this, assuming that you haven't.]

  5. William said,

    May 20, 2011 @ 3:30 pm

    @secretseasons: I think you're confused about the terminology here. Speakers may use different phones (physical sounds) to express the same phoneme (distinctive sounds of a language).

    In other words, speakers of different dialects of English demonstrate phonetic diversity–that's why they sound different–but usually not phonemic diversity–that's why they're mutually intelligible.

  6. KWillets said,

    May 20, 2011 @ 3:31 pm

    I wonder if this might be resolved by measuring the information content per phoneme rather than the mere number of phonemes. If a language truly has more useful phonemes, then its average word or sentence length will be lower.

    [(myl) See e.g. this, this and this. Unfortunately, WALS does not even count "number of phonemes", much less consider their redundancy in context, but rather bins tones, vowels, and consonants separately into three classes each. Atkinson then averages the three numbers to get his measure of phonemic diversity.]

  7. Keith M Ellis said,

    May 20, 2011 @ 5:42 pm

    I'm quite strongly in favor of interdisciplinary work; but what the hell is it with these non-linguists doing historical linguistics simply because they think their familiar models and mathematical tools will apply? It seems this happens with anthropology, too, with molecular biologists doing population migration stuff in almost complete ignorance of all the preexisting work in the field.

    [(myl) I think your objection is unfair in this case. The basic issues and the basic mathematics of phylogenetic reconstruction, for example, are pretty much the same in historical biology and in historical linguistics, even if the traits, the taxonomy of changes, the guiding forces etc. are different. There's certainly room to argue about which algorithms to use -- in both kinds of reconstruction. And in this case, Atkinson's measure of "diversity" strikes me as a very problematic one, for reasons that I explained at greater length here. But there's nothing wrong with trying.]

    Atkinson is affiliated with psychology and anthropology departments and presumably he worked with some linguists (one hopes). But the NYT article describes him as a biologist…which isn't the least surprising, is it?

    I'm not thrilled with a lot of the work in the social sciences–often, much of it seems too ideologically-determined to me. Nevertheless, there is a burgeoning and pernicious problem with non-social scientists taking it upon themselves to do work in the social sciences in complete disregard of what's already there and in complete confidence that they're the "real" scientists who are going to finally bring rigor to the matter.

    I never cease to be amazed at how the same scientist who knows so intimately that someone in the same field but a different subspecialty is basically clueless can then turn around and speak, write, and even attempt to work authoritatively in an entirely different discipline! The mind boggles! Or my mind boggles, anyway.

  8. John Roth said,

    May 20, 2011 @ 7:02 pm

    @Keith

    I'm somewhat surprised it got published in Science because the serial founder effect for human dispersal is quite well established on a genetic basis. Even if this paper was correct, all it does is document that an already well known effect also shows up in a different domain.

    As far as anthropological migration work, the last couple of years have seen an almost complete reversal of the conventional picture of European settlement by agriculturists at the expense of previous hunter-gatherers, and it's all based on the molecular biology. The whole question of what really happened is still up in the air, although it looks like the diffusionists are in panicked retreat.

    If you want to get to the bottom of many fields these days, you almost have to do multi-disciplinary work.

  9. Simon Greenhill said,

    May 20, 2011 @ 7:12 pm

    "what the hell is it with these non-linguists doing historical linguistics simply because they think their familiar models and mathematical tools will apply?"

    As one of these "non-linguists" AND one of those "non-anthropologists", here's why: The methods have been designed to solve the same problems. Especially when it comes to historical linguistics.

    There've been a number of papers justifying and explaining this – I've written at least two papers on this and Quentin Atkinson has written some too.

    Simon

    [(myl) I entirely agree with you about the general approaches, and I mostly agree with your about the analytical details in many cases (for what little this is worth, since I'm not an expert in most of the areas involved). I do know something about phonology, phonetics, the WALS survey, and the algorithms involved in this particular discussion; and I feel that A's way of measuring phonological diversity is really not a good one, and that as a result, the observed effect is probably an artefact of areal features, just as Goddard and Frohlich suggest. On the other hand, I think that the "what the hell is it with these non-linguists..." attitude is completely inappropriate. If this paper is wrong, it's not because Prof. Atkinson lacks the right union card or guild membership, but because he accepted the (in my opinion rather bad) decision (made by the card-carrying linguist Ian Maddieson) to make a coarse count of categories that doesn't consider syllable-level context effects; and compounded this problem by using an unweighted average of tone, vowel, and consonant "diversity". ]

  10. Zach Blume said,

    May 20, 2011 @ 7:20 pm

    A few questions :)

    1.) Following up on what secretseason has said — maybe I'm misunderstanding your rebuttal slightly — why doesn't the existence of phonemic mergers point to the validity of the genetic drift/"founder's effect" argument? Theoretically, if a small group with a phoneme merger doomed to be erased by random processes split off from the general linguistic community right now speaking English into some unknown frontier, I would think it is plausible to wager that their chances of preserving this phonemic merger is higher?

    2.) In the quote you published, "no mechanism of language change is known that would plausibly explain why speakers should lose phonemes from a language they already speak just because the speech community becomes smaller"

    Shouldn't the general concept of spreading unmarked features through a speech community cover this? It is more marked, in some sense, to create a new phoneme than to have a merger — no? If phonemic mergers are the dominant trend in phonemic innovation, rather than the creation of/splitting into new phonemes, then a random sampling effect like genetic drift would imply a random sample has a higher chance of selecting a group of individuals with less distinctions?

    3.) On another note, as far as how Atkinson "binned" the phoneme diversity (few, medium, lots) seems to me rather reminiscent of an ecological fallacy, though I am not thoroughly versed in statistics to really know. Does Atkinson check to see that the unbinned values correlate just as well?

    Thanks!

  11. Ran Ari-Gur said,

    May 20, 2011 @ 7:23 pm

    @Keith M Ellis: obligatory xkcd link

  12. Xmun said,

    May 20, 2011 @ 8:59 pm

    @John Cowan: If you must quote Jakob Grimm in the original German, it would be a kindness to give the author's name — even here, it shouldn't be assumed that we all know — and to offer a translation. And is your text correct? I've seen it rendered with "hochmütiger, vornehmer, streitsüchtiger" rather than "stolzer, edler, streitsüchtiger".

  13. Chad Nilep said,

    May 20, 2011 @ 9:11 pm

    "If you want to get to the bottom of many fields these days, you almost have to do multi-disciplinary work."

    I quite agree, but multi-disciplinary work should be, um, multi-disciplinary. To me, that implies engaging with existing theory in field A and using theory, data, or methods from field B to problematize/strengthen/reject it. What is too often missing is the first step – engaging with existing theory.

    "The methods have been designed to solve the same problems."

    In fact I would say that the methods have been designed to solve problems that appear similar. This is much like arguing from analogy: often it works well and leads to new insights, but only to the extent that the analogy is apt. Phonemes are not genes, languages are not species; they may appear to evolve similarly in some respects, but they differ in other respects.

  14. Simon Greenhill said,

    May 21, 2011 @ 12:22 am

    @Chad – Of course. It's absolutely about taking a fairly general set of powerful tools for testing *evolutionary* hypotheses and adapting them to answer questions about languages and cultures. It's not about trying to jam a square data peg into a round methods hole.

  15. Richard Sproat said,

    May 21, 2011 @ 11:19 am

    I agree with Goddard and Frohlich's argument. Indeed I made essentially the same
    point in my review of the paper for Science. Since they chose to ignore the
    point then, it's hardly surprising that they would choose to ignore a letter
    making the same point. Science does not seem to provide an open forum for
    experts to provide serious critiques of "groundbreaking" papers. We had the same
    experience when we tried to publish a critical letter of the Rao et al. work on
    the Indus symbols back in 2009.

    But there's something more interesting. One of my other critiques is that while
    Atkinson (and before him Hay & Bauer) has shown that there is some sort of
    correlation between population size and phonemic diversity for modern speaker
    population ranges, he did not demonstrate any such correlation for ranges
    plausible for the Paleolithic. Whereas the modern correlation is a modest
    r=0.385, just looking at his plots, if one restricted the range to a few tens of
    speakers to a few thousand, it did not seem likely that there would be a
    correlation.

    As it turns out, Atkinson has addressed this point, sort of. In his online
    supplement we find the following addition to his caption for Figure S1:

    "The same relationship holds when the analysis is restricted to languages
    with speaker populations of 5000 or less, a range in line with speaker
    populations of modern hunter-gatherers (S5) and likely to have characterised
    much of human prehistory." [Online supplement, Page 12]

    He goes on to claim a p value of well under .001. The r value he reports is
    r=-0.250.

    Am I missing something? Or is this a bald-faced attempt to slip something over
    on the reader? That is a *negative* correlation: so over the population ranges
    in question, the smaller the population, the higher the phonemic diversity.

  16. Keith M Ellis said,

    May 21, 2011 @ 12:58 pm

    @Simon, you wrote:

    "Of course. It's absolutely about taking a fairly general set of powerful tools for testing *evolutionary* hypotheses and adapting them to answer questions about languages and cultures. It's not about trying to jam a square data peg into a round methods hole."

    But you've ignored the Chad's essential point:

    "To me, that implies engaging with existing theory in field A and using theory, data, or methods from field B to problematize/strengthen/reject it. What is too often missing is the first step – engaging with existing theory."

    Anyone defending such work will always say that the tools they're using are appropriate to the work they're doing. They wouldn't be doing the work or defending it otherwise. But because they believe this doesn't make it true.

    BTW, you might have mentioned that you've collaborated with Atkinson for years.

    What I think Chad is saying is that too often people will apply a familiar approach from their field to a problem in another field, producing results that are inconsistent with existing work in that field without engaging, or often even being familiar with, such work.
    @Mark wrote:

    "On the other hand, I think that the "what the hell is it with these non-linguists…" attitude is completely inappropriate. If this paper is wrong, it's not because Prof. Atkinson lacks the right union card or guild membership, but because he accepted the (in my opinion rather bad) decision (made by the card-carrying linguist Ian Maddieson) to make a coarse count of categories that doesn't consider syllable-level context effects; and compounded this problem by using an unweighted average of tone, vowel, and consonant "diversity". "

    You've made my point for me, really. I think you realize that I'm not asserting that credentials or "guild membership" is the problem per se, but rather indicative or suggestive of a problem of competency. This was Atkinson's work and paper, he bears the ultimate responsibility for relying upon Maddieson.

    I don't have the competency to evaluate the technical issues. Perhaps this particular research, whatever its problems, is not an example of the problem I'm objecting to. I'd be happy to know that. That this is the case, though, doesn't invalidate my general complaint.

    In any case, however, given your criticism, it certainly sounds as if there were some errors made related to basic competency in the very linguistics-specific attributes of the data.

    Quoting you from a post you wrote years ago about work by Atkinson and Gray:

    "Their (very brief) paper and their bibliography make it clear that they are quite familiar with the relevant literature and have in some way taken account of these issues in their modeling."

    …which is the competency that I am asserting ought to exist, but far too frequently does not exist in similar cross-disciplinary work.

    As an informed layperson, it's my strong impression that I've seen a surprising number of widely discussed papers by, in particular, physicists and biologists, on surprising results in the social sciences that, when evaluated by social scientists, show a remarkable ignorance of pre-existence work and results and, quite often, are deeply flawed as a result. But the work is widely reported, presumably cited, and likely more physicists and biologists come to believe that their pet mathematical models are appropriate for doing ground-breaking work in whatever subject strikes their fancy.

    Please keep in mind that as a layperson, I have no professional stake in these matters; I'm not defending my turf. Indeed, due to some of my educational background and, more importantly, my own temperaments and interests, if I do have an allegiance, it would almost certainly be with…the physicists and biologists and not the social scientists. And, if I were a researcher, I'd almost certainly want to be working in a multi-disciplinary environment. Maybe I'm hyper-correcting my own biases.

    This all brings to mind some of the related debate that's taken place elsewhere in the blogosphere, particularly the cross-blog discussion that originated on Crooked Timber in 2005.

    There's no doubt that there's important similarities between all sorts of diverse natural phenomena, including within the domain of the social sciences. I am a long, long enthusiast of complexity theory, for example. But that's also a counter-example because there's been a fair bit of crap and overly-ambitious and hubristic work done in complexity to which my general criticisms here apply. If one is going to attempt to do this kind of cross-disciplinary work, then one has a responsibility to arm oneself adequately with both substantial familiarity/competency in the target specialty, and also with humility. The most egregious work results from researchers who have little respect for the existing work in the target field.

  17. Dragos said,

    May 21, 2011 @ 2:59 pm

    Speaking of missing first steps, there's one other reason I find many phylogenetic reconstructions unconvincing. I think a valid (scientific) model must predict (=confirm) the things we know, before using it to predict the things we don't know. Reading only several such papers, I have not seen a "New phylogenetic tree model correctly predicts the formation of the Romance languages" (or Slavic, or whatever well-studied language group)

  18. Tom D said,

    May 21, 2011 @ 3:44 pm

    @Dragos

    Speaking of missing first steps, there's one other reason I find many phylogenetic reconstructions unconvincing. I think a valid (scientific) model must predict (=confirm) the things we know, before using it to predict the things we don't know. Reading only several such papers, I have not seen a "New phylogenetic tree model correctly predicts the formation of the Romance languages" (or Slavic, or whatever well-studied language group)

    There are a good number of these types of papers. The most recent one I know of was by Lee and Hasegawa (2003) on Japonic. Others include ones on Austronesian (Greenhill and Gray (2009)), Indo-European (Gray and Atkinson (2003) and Warnow et al. (2004), among many others), and Bantu (Holden and Gray (2006)).

  19. Dragos said,

    May 21, 2011 @ 6:31 pm

    @Tom D

    There are a good number of these types of papers. The most recent one I know of was by Lee and Hasegawa (2003) on Japonic. Others include ones on Austronesian (Greenhill and Gray (2009)), Indo-European (Gray and Atkinson (2003) and Warnow et al. (2004), among many others), and Bantu (Holden and Gray (2006)).

    Some of these papers I've read and this is exactly the approach I find unscientific and unconvincing. If these models can't predict the split of Vulgar Latin or Common Slavic into their daughter languages, then why would they work for the proto-languages we obviously don't know? Mutatis mutandis, this question should be asked in the quest for the origin of language, regardless of method and model.

  20. Simon Greenhill said,

    May 21, 2011 @ 8:23 pm

    @Dragos – ok, then, how about this then? I built a phylogeny of 400 Austronesian languages. Of the 34 major subgroups listed as well-established by Bob Blust, we got 25. And the ones we didn't get are either language linkages rather subgroups, or represent potential subgrouping hypotheses. Overwhelmingly the phylogenetic results were very close to the historical linguistics work, and far far closer to this than chance. I discussed this in great detail here..

  21. Simon Greenhill said,

    May 21, 2011 @ 8:35 pm

    @Keith.

    First of all – I'm absolutely not hiding my collaborations with Quentin. We've done some very cool work together, and we're still working on cool things together.

    In terms of engaging with linguists (and anthropologists): I absolutely see the need for this and I do this every day. I have spent a hell of a lot of time engaging with linguists. I've been to more linguistics/anthropology conferences then I have any other conferences. I go to linguistics and anthropology departments and give talks. I build resources for linguists ([1], [2]). I work closely with linguists and have published with them. I've got a paper coming out in a linguistics journal in the next few months. I read a hell of a lot of papers in these fields too. I think pretty much every single paper I've ever published I've sent to at least two linguists and asked them for comments. Hell, one of the reviewers of my PhD was a linguist.

    I probably shouldn't speak for Quentin, or anyone else, but I know he feels the same way.

    Personally – I'd like to see some linguists engaging with this work more, rather than just dismissing it immediately because it's "not linguistics". This is a really shallow view of science. The questions are important not which department you're in.

    Simon

  22. Keith M Ellis said,

    May 21, 2011 @ 9:24 pm

    "Personally – I'd like to see some linguists engaging with this work more, rather than just dismissing it immediately because it's "not linguistics". This is a really shallow view of science. The questions are important not which department you're in."

    I strongly agree.

    Let's shift the context a bit. Take the example of evolutionary psychology. There's a lot of crap work being done in EP by ideologically biased and arrogant researchers who quite intentionally ignore prior work because they believe they have the right tools and the right models where, in contrast, those who came before were hopelessly confused. That's an exaggeration (I'd like to believe), but it's not as far off-the-mark as we might wish.

    As it happens, I think that there is quality work being done in EP and the field does have something important to offer. It's just that there's been a critical mass of crap work that has stained the reputation of the entire field and, ultimately, has gotten in the way of doing productive science.

    My point is that I applaud your care to avoid the kinds of mistakes I'm describing. My complaints are not in any way intended to imply that there's no good and important work being done in linguists (or anthropology) by biologists and others. It's that there's, as far as I can tell, too much not good work and that's not only a bad thing in itself, but it also gets in the way of the good work that's being done, as well. And this all would be avoided if people who work on topics outside their specialty made certain that humility and diligence were core values while doing so.

  23. John Cowan said,

    May 22, 2011 @ 12:32 am

    Xmun: I don't expect you to recognize the quotation, but I expect you can google it. I don't guarantee that it came straight from Grimm's pen to my keyboard unchanged.

    myl: Wells's lexical sets, though a wonderful invention, are consciously restricted to only two varieties of English, RP and "General American" (that's horror quotes, for there really is no such thing any more). The set I'm using now, which covers (I hope) all the living varieties of English speech, is Wells/Mills/Cowan/Rosta:

    KIT, DRESS, TRAP, BAD, LOT, STRUT, FOOT, BATH, DANCE, CLOTH, NURSE, TERM, DIRT, FLEECE, BEAM, FACE, TRAIL, FREIGHT, PALM, THOUGHT, GOAT, SNOW, GOOSE, THREW, PRICE, CHOICE, MOUTH, NEAR, SQUARE, START, NORTH, FORCE, CURE.

    And for the fun of it, my personal mergers are TRAP=BAD=BATH=DANCE, LOT=PALM, CLOTH=THOUGHT, NURSE=TERM=DIRT, FLEECE=BEAM, FACE=TRAIL=FREIGHT, GOAT=SNOW, GOOSE=THREW, NORTH=FORCE.

  24. Dragos said,

    May 22, 2011 @ 11:43 am

    @Simon, I'm no linguist and I do not know much about Austronesian languages. In my opinion, a phylogenetic tree model should be used especially for remote Oceania, because of the geography. Clear splits, in many cases no further contact, it seems easy enough to model using a tree. But is this the case for all language groups? For all Austronesian languages?

    When reconstructing proto-worlds, I cannot agree with the postulate equating language with material culture. I'm quoting from William Foley's The Papuan languages of New Guinea (where also some Austronesian languages are spoken), p. 26: "there is often no correlation between cultural patterns and the linguistic affilitations of communities. For example, the southernmost groups of Arapesh speakers are culturally very similar to the adjoining Abelam group, in spite of the fact that the languages are in structure radically different and completely unrelated. On the other hand, while the language of these southern Arapesh is very similar to that of the Mountain Arapesh described by Mead, the cultures are strikingly different. Communities must thus be viewed as foci in areal networks of cultural and linguistic patterns."

    Moreover in New Guinea there's much linguistic diversity. Multilingualism is (and was) quite common. "Other indigenous languages which functioned as lingue franche in pre-contact times were the Austronesian languages, Motu of the cental Papuan coast, Suau of the southwest tail area of Papupa, and Dobu of the islands of eastern Papua" (p. 31) Doesn't that mean also a different dynamics of language contact and shift? Among the languages spoken here, even basic vocabulary was extensively borrowed (e.g. from Austronesian languages to Papuan languages). With no written records from the past, historical linguists themselves can't reliably reconstruct high-level language groups and families.

    Therefore assumptions like "languages correlate with archaeological cultures" (read also Paul Heggerty's Linguistics for Archaeologists: Principles, Methods and the Case of the Incas) or "basic vocabulary is resistant to borrowing" are not granted for every language, not for every proto-language.

    Overwhelmingly the phylogenetic results were very close to the historical linguistics work, and far far closer to this than chance.

    Let me explain my skepticism using an analogy. There was once a widely held Ptolemaic model, very close to the observations of the movements of celestial bodies, and presumably many of its proponents believed it was "far far closer to reality than chance".

  25. Keith M Ellis said,

    May 22, 2011 @ 1:45 pm

    @Dargos,

    Let me explain my skepticism using an analogy. There was once a widely held Ptolemaic model, very close to the observations of the movements of celestial bodies, and presumably many of its proponents believed it was "far far closer to reality than chance".

    I seem to have a hobby defending Ptolemaic astronomy. I don't think that your example proves the point as convincingly as you'd like it to.

    Until Galileo's telescopes, Ptolematic astronomy perfectly agreed with observation and was perfectly predictive. This was about a century after Copernicus. The heliocentric model was known to Ptolemy and the ancients, and they were well aware of its geometric simplicity relative to the geocentric model. However, heliocentricism required the assumption that the Earth was in motion, and implied that the cosmos was unimaginably vast. Both were quite justifiably difficult or impossible to accept. The oft-maligned epicycles were no more absurd than normalizing variables in equations today. More problematic, in my personal opinion, was the equant–but, really, that's quibbling. So, all in all, Ptolemaic astronomy was a pretty good theory. The important thing is that it conformed perfectly to observation and conformed well with the preexisting theories of the nature of the cosmos.

    Indeed, one can make the argument that until Galileo–that is, until there was actual evidence that the geocentric model couldn't explain observation–it was Copernicus who was arrogant and premature. I wouldn't actually make that argument, however, because in contrast to my complaints detailed earlier, Copernicus was well-versed in geocentricism and all the available data and was, in fact, greatly elaborating a pre-existing notion of heliocentricism in an extremely rigorous and knowledgeable fashion.

    Even when there's a paradigmatic revolution, the new theories always engage discursively with the old theories. That's why I think that the deeper problem here (not necessarily in the particular example of the subject of this post) is cross-disciplinary work that's done with too much arrogance and too little familiarity with preexisting work in the target discipline. The problem is not with bringing a fresh view and new tools to a problem. That's good.

  26. Richard Sproat said,

    May 22, 2011 @ 2:14 pm

    BTW, I'm curious: did Science decline to publish the Goddard/Frohlich letter for reasons related to "lack of space"?

    That was their stated reason for our letter relating to the Rao et al. paper in 2009. Since this is all electronic, that struck me as a lame reason.

  27. Dragos said,

    May 22, 2011 @ 4:24 pm

    @Simon,

    I don't think the analogy holds to that extent. As you well noted, the predictions of Ptolemaic astronomy agreed with the observations (but not perfectly). However these phylogenetic trees predict the unobservable (and the fit is far from being perfect). I can't say we have a pretty good theory, because it's not empirical science (yet). I think it can be (and I certainly do agree that bringing fresh views and tools is good). It's just that at this moment it seems more important to arbitrate between competing theories and reconstructions of proto-worlds and proto-languages, instead of validating such models against observed data.

  28. Dragos said,

    May 22, 2011 @ 6:24 pm

    I apologize for addressing my previous reply to Simon, instead of Keith.

  29. Simon Greenhill said,

    May 22, 2011 @ 8:25 pm

    It's easy to point out examples where languages and cultures and genetics don't mesh nicely. Melanesia is, of course, one of the prime examples. However, the critical issue is how likely is it that there's a complete decoupling of these lines of evidence? Sure, languages and cultures borrow things from each other, sure gene flow happens, but what level of diffusion do you need to postulate to have the story of the genes, the languages, and the cultures saying completely different things? What process would cause complete decoupling? There's going to be a continuum here from completely decoupled to tightly linked. Where on that continuum will societies land? that's going to be an empirical question.

    I'd also like to point out a simulation study I did here showing that phylogenetic methods are actually quite robust to the levels of borrowing we see in most languages. I've also co-authored a study here showing that comparative phylogenetic methods are powerful ways of inferring ancestral states regardless of borrowing.

    In addition – studies that have looked at the coherence between gene/language/cultural lineages around the world have usually shown high levels of concordance, including in Bougainville (Howells 1966), Venezuela (Spielman et al. 1974), Europe (Sokal 1988, Barbujani and Sokal 1990). There are strong correlations in Europe & East/Central Asia, but lower concordance in West Africa and South-East Asia (Nettle and Hariss 2003). There's a strong match between genes, languages and Y chromosome lineages on Sumba (Lansing et al 2007). And there's quite a bit of work been done in Melanesia showing that it's not as bad as everyone makes out (e.g. Hunley et al 2008).

    Simon

  30. Ives Goddard said,

    May 23, 2011 @ 8:41 am

    To answer Richard Sproat's question:

    The boilerplate rejection letter from Science said:
    "We receive many more letters than we can accommodate and so we must reject most of those contributed. Our decision is not necessarily a reflection of the quality of your letter but rather of our stringent space limitations."

    This claim was immediately belied by an invitation "to resubmit [our] comments as a Technical Comment" of "up to 1000 words in length."

    Perhaps someone else's letter on the Atkinson article will be more successful, but we do not feel encouraged to try to come up with a critique that would both point out the evident flaws in the original article (to the list of which Sproat now adds another serious one) and pass muster with the editors.

  31. chris said,

    May 23, 2011 @ 9:50 am

    It's easy to point out examples where languages and cultures and genetics don't mesh nicely. Melanesia is, of course, one of the prime examples.

    The first ones that came to my mind were African-Americans and Jews. Different groups of Jews genetically resemble the populations they dwelt among for centuries, but their language and culture remains more like each other; African-Americans are just the opposite, with largely African genetics but (in most cases) speaking a Germanic language and with a culture that is (in substantial part) neither African nor European, but shaped strongly by their more recent history.

    Central and South Americans are another good example (or several). There must be millions of Spanish- and Portuguese-speaking Catholics with little or no genetic connection to Rome.

    And then there's English speakers (or for that matter Muslims) in India… I could go on, but I think you get the point. Many examples from the last thousand years are well-known and well-documented, so it may be tempting to overlook them, but how many similar events are buried deeper in history?

    However, the critical issue is how likely is it that there's a complete decoupling of these lines of evidence? Sure, languages and cultures borrow things from each other, sure gene flow happens, but what level of diffusion do you need to postulate to have the story of the genes, the languages, and the cultures saying completely different things? What process would cause complete decoupling?

    In the above cases of complete or near-complete decoupling, it seems to often be not a matter of diffusion or borrowing, but of large, often violent, historical events.

  32. Richard Sproat said,

    May 23, 2011 @ 3:46 pm

    @Ives Goddard

    Thanks.

    They evidently have something that generates that verbiage. This is the identical reply that we had received back in 2009.

  33. Dragos said,

    May 25, 2011 @ 8:14 pm

    @Simon

    In addition – studies that have looked at the coherence between gene/language/cultural lineages around the world have usually shown high levels of concordance, including in Bougainville (Howells 1966), Venezuela (Spielman et al. 1974), Europe (Sokal 1988, Barbujani and Sokal 1990). There are strong correlations in Europe & East/Central Asia, but lower concordance in West Africa and South-East Asia (Nettle and Hariss 2003). There's a strong match between genes, languages and Y chromosome lineages on Sumba (Lansing et al 2007). And there's quite a bit of work been done in Melanesia showing that it's not as bad as everyone makes out (e.g. Hunley et al 2008).

    … studies showing languages may correlate with genes, not with material culture. That "Swede metaphor" is so 19th century.
    As for genetics, there are too many exceptions (looking at the genetic trees based on distance by Cavalli-Sforza et al, French is grouped with Germanic populations but Icelandic is not, Hungarian is grouped with Polish, etc.) and often such correlations can be rather justified by geographic proximity. For a more detailed discussion see ch. 5 "Correlations between genetic and linguistic data" in April McMahon & Robert McMahon, Language Classification by Numbers.

    I'd also like to point out a simulation study I did here showing that phylogenetic methods are actually quite robust to the levels of borrowing we see in most languages. I've also co-authored a study here showing that comparative phylogenetic methods are powerful ways of inferring ancestral states regardless of borrowing.

    Robust is a two-edged word. Looking at the series of studies by R. Gray, Q. Atkinson, G. Nicholls et al. on the IE language tree, one can say their methods are robust. But together with their estimate (~8700 years BP), there are also some errors occuring in most or all trees. I'll just focus on the Romance languages. For instance, placing Catalan and Provençal in two different clades is dubious and so is grouping Romanian with Ladin (and in the latter case significant borrowings can be safely excluded).

    I guess such errors in topology are (also) caused by differences between models and reality. In Heggarty's paper above you'll find most of these languages described in a wave model (dialect continuum), instead of a tree model (read also this). Some cognate sets were not formed on separate branches, but in the regional varieties of the root language. Center/periphery isoglosses may be responsible for cognate sets separating Romanian and Ladin from the rest. West/east isoglosses (like the cognate sets derived from plorare and plangere) may be responsible for a Gallo-Iberian group. Certainly there were some Romance regional developments as well. Topology is not necessarily phylogeny, and branch lengths may be misleading. A simple tree with Italian, Spanish, French would probably be best drawn with three branches stemming from one root, if phylogeny is what we have in mind.

    I also played a little with TraitLab. I used Gray's matrix with 2665 characters for 16 Romance languages. I ran the program ~20 times. I used different sets of calibrations (pre-defined clades) and I obtained root dates between 1900 and 2600 years BP (in Gray's IE data file this node was dated between 1700 and 1850 years BP for calibration). When using only recent information (with two dated clades: French creoles and French, and Brazilian and Portuguese), there were many unfactual dates (the splits between Spanish and Portuguese, and French and Provençal were dated between 600 and 900 years BP), and the root was variously dated between 1900 and 2300 years BP. By adding more clades (e.g. grouping Catalan and Provençal) and dates (e.g. we have Old French attested 1200 years BP), I pushed the root date further back in time. Apparently we can't do dates if we lack contemporary historical information.

  34. Simon Greenhill said,

    May 27, 2011 @ 4:39 am

    @Dragos.

    I'm travelling at the moment so it's hard to get internet access. Apologies, etc.

    1. As I said – there are are examples of where there is discontinuity between genes, languages, and cultures. The issue is how common this is, and what _process_ do you need to explain this. The information from the genes, the languages, and the cultures are being drawn from the same underlying source – human prehistory – and they should therefore show strong concordance unless there's something unusual happening.

    Also note that this is nothing to do with language phylogenies, per se, but a general issue with historical linguistics, archaeology, and genetics. Denying that these sources provide information about the pattern of human prehistory is utterly nihilistic.

    2. As an aside – the networks Heggarty uses (in various papers, which I am very aware of) are not wave models. These are phylogenetic algorithms first and foremost. For whatever that's worth.

    3. As for the Romance languages issue. There may be some small glitches BUT First and foremost this is a statistical estimation problem. Phylogeny is hard. When there are, say 9 languages there are 2 million possible trees. Add one more language, and that increases to 34 million possibilities.

    It is absolutely not realistic to expect that with finite data and with finite time you're going to resolve everything – instead, you take an estimate using the data you have and you quantify the robustness of that estimate. Catalan and Provençal, for example, in the Gray and Atkinson trees have low posterior probabilities (57% and 83%). This means that their locations are not very well estimated – this could be due to lack of signal in the data they used, it could be a problem with overlapping isoglosses or borrowing or whatever. I would point out that these measures of strength-of-evidence go way beyond what a family tree is in historical linguistics. To really understand what's going on there you need to look at the entire distribution of trees.

    More importantly this topological error is only relevant if it impacts the hypothesis being tested. If the Gray and Atkinson paper was concerned with the subgroupings of Romance languages then this would be crucial. However, they're not – they're estimated the age of the family. I've shown in some of the publications I linked above that these estimates are highly robust to minor topological mis-specifications.

    4. Re: Traitlab – That's interesting. Some ideas – Traitlab (or more accurately, the stochastic dollo model that it uses) does not handle borrowing very well at all (i.e. it assumes one origin event for each cognate). If there is unidentified borrowing in the Romance languages then this will cause model misspecification – probably pushing that root age back deeper. The binary models in MrBayes (which G&A used in their 03 paper) or BEAST are more robust to these effects. Second, dating is difficult – if you only have shallow calibrations then the deeper branches will be poorly specified.

    Simon

  35. Rory Van Tuyl said,

    February 9, 2012 @ 6:28 pm

    Science Magazine has, at long last, published three "Technical Comments" in response to Atkinson's original article. All are unstinting in their criticism of his paper "“Phonemic Diversity Supports a Serial Founder Effect Model of Language Expansion from Africa” Science 15 April 2011, p. 346.

    To view these detailed critiques, go to:
    http://www.sciencemag.org/content/current#TechnicalComments

RSS feed for comments on this post

Leave a Comment