Tuesday, April 19, 2011

Universal blues

I have written a news story and a leader for Nature on a new paper examining the notion that there are universal grammatical principles in language. Here they are, in that order. But I must say that, much as the results reported by Dunn et al. chime with my instinctive resistance to universal theories of anything, the comments I’ve received on the paper make me a little sceptical that it does what it claims. Time will tell, I suppose.

Linguists debate whether languages share universal grammatical features.

Languages evolve in their own idiosyncratic fashion, rather than being governed by universal rules. That’s the conclusion of a new study which compares the grammar of several hundred languages in the light of their evolutionary trees.

Psychologist Russell Gray of the University of Auckland in New Zealand and his coworkers examine the relationships between traits such as the ordering of verbs and nouns in four families representing more than 2,000 languages, and find no sign of any persistent, universal guiding principles [1].

It’s already proving to be a controversial claim. “There is nothing in the paper that brings into question the views that they are arguing against”, says linguist Matthew Dryer of the State University of New York at Buffalo.

There is thought to be around 7,000 languages in the world, which show tremendous diversity in structure. Some have complex ways of making composite words (such as Finnish), others have simple, short and invariant words (such as Mandarin Chinese). Some put verbs first in a sentence, others in the middle and others at the end.

But many linguists suspect there be some universal logic behind this bewildering variety – common cognitive factors that underpin grammatical structures. Two of the most prominent ‘universalist’ theories of language have been proposed by American linguists Noam Chomsky and Joseph Greenberg.

Chomsky tried to account for the astonishing rapidity with which children assimilate complicated and subtle grammatical rules by supposing that we are all born with an innate capacity for language, presumably housed in brain modules specialized for language. He suggested that this makes children able to generalize the grammatical principles of their native tongue from a small set of ‘generative rules’.

Chomsky supposed that languages change and evolve when children reset the parameters of these rules. A single change should induce switches in several related traits in the language.

Greenberg took a more empirical approach, enumerating many observed shared traits between languages. Many of these concerned word order. For example, a conditional clause normally precedes its conclusion: “if he’s right, he’ll be famous.” Greenberg argued that these universals reflect fundamental biases, probably for cognitive reasons. “The Greenbergian word order universals have the strongest claim to empirical validity of any universalist claim about language”, says Gray’s coauthor Michael Dunn of the Max Planck Institute for Psycholinguistics at Nijmegen.

Both of these ideas have implications for the family tree of language evolution. In Chomsky’s case, as languages evolve, certain features should co-vary because they are products of the same underlying parameter. Greenberg’s idea also implies co-dependencies between certain grammatical features of a language but not others. For example, the word order for verb-subject pairs shouldn’t depend on that for object-verb pairs.

To test these predictions, Gray and colleagues used the methods of phylogenetic analysis developed for evolutionary biology to reconstruct four family trees representative of more than 2,000 languages: Austronesian, Indo-European, Bantu and Uto-Aztecan. For each family they looked at eight word-order features and used statistical methods to calculate the changes that each pair of features had evolved independently or in a correlated way. This allowed them to deduce the webs of co-dependence among the features and compare them to what the theories of Chomsky and Greenberg predict.

They found that neither of these two models matched the evidence. Not only do the co-dependencies differ from those expected from Greenberg’s word-order ‘universals’, but they are different for each family. In other words, the deep grammatical structure of each family is different from that of each of the others: each family has evolved its own rules, so there is no reason to suppose that these are governed by universal cognitive factors.

What’s more, even when a particular co-dependency of traits was shared by two families, the researchers could show that it came about in different ways for each – that the commonality may be coincidental. They conclude that the languages – at least in their word-order grammar – have been shaped in culture-specific ways and not by universals.

Other experts express some skepticism about the new results, albeit for rather different reasons. Martin Haspelmath at the Max Planck Institute for Evolutionary Anthropology in Leipzig, Germany, says he agrees with the conclusions but that “for specialists they are nothing new”. “It’s long been known that grammatical properties and dependencies are lineage-specific”, he says.

Meanwhile, Dryer, who has previously presented evidence that supports Greenberg’s position, is not persuaded that the results make a convincing case. “There are over a hundred language families that the authors ignore but which provide strong support for the views they are arguing against”, he says. There is no reason to expect a consistent pattern of word-order relationships within families, he adds, regardless of whether they are shaped by universal constraints.

Haspelmath feels it may be more valuable to look for what languages share in common than how they (inevitably) differ. Even if cultural evolution is the primary factor in shaping them, he says, “it would be very hard to deny that cognitive biases play no role at all.”

“Comparative linguists have focused on the universals and cognitive explanations because they wanted to explain something”, he adds. “Saying that cultural evolution is at play basically means that we can’t explain why languages are the way they are – which is largely true, but it’s not the whole truth.”

1. Dunn, M., Greenhill, S. J., Levinson, S. C & Gray, R. D. Nature 10.1038/nature089923 (2011).

A search for universals has characterized the scientific enterprise at least since Aristotle. In some ways, this quest for common principles underlying the diversity of the universe defines science: without it there is no order and pattern, but merely as many explanations as there are things in the world. Newton’s laws of motion, the oxygen theory of combustion and Darwinian evolution each united a host of different phenomena in a single explicatory framework.

One view takes this impulse for unification to its extreme: to find a Theory of Everything that offers a single generative equation for all we see. It is becoming ever less clear, however, that such a theory – if it exists – can be considered a simplification, given the proliferation of dimensions and universes it might entail. Nonetheless, unification of sorts remains a major goal.

This tendency in the natural sciences has long been evident in the social sciences too. Darwinism seems to offer justification: if all humans share common origins, it seems reasonable to suppose that cultural diversity must also be traceable to more constrained origins. Just as the bewildering variety of courtship rituals might all be considered forms of sexual selection, so perhaps the world’s languages, music, social and religious customs and even history could be governed by universal features. Filtering out what is contingent and unique from what is shared in common might enable us to understand how complex cultural behaviours arose and what ultimately guides them in evolutionary or cognitive terms.

That, at least, is the hope. But a comparative study of linguistic traits by Dunn et al. (online publication doi:10.1038/nature09923) supplies a sharp reality check on efforts to find universality in the global spectrum of languages. The most famous of these was initiated by Noam Chomsky, who postulated that humans are born with an innate language-acquisition capacity – a brain module or modules specialized for language – that dictates a universal grammar. Just a few generative rules are then sufficient to unfold the entire fundamental structure of a language, which is why children can learn it so quickly. Languages would diversify through changes to the ‘parameter settings’ of the generative rules.

In contrast, Joseph Greenberg took a more empirical approach to universality, identifying a long list of traits (particularly in word order) shared by many languages, which are considered to represent biases that result from cognitive constraints. Chomsky’s and Greenberg’s are not by any means the only theories on the table for how languages evolve, but they make the strongest predictions about universals. Dunn et al. have put them to the test by using phylogenetic methods to examine the four family trees that between them represent over 2,000 languages. A generative grammar should show patterns of language change that are independent of the family tree or the pathway tracked through it, while Greenbergian universality predicts strong co-dependencies between particular types of word-order relations (and not others). Neither of these patterns is borne out by the analysis, suggesting that the structures of the languages are lineage-specific and not governed by universals.

This doesn’t mean that cognitive constraints are irrelevant, nor that there are no other universals dictated by communication efficiency. It’s surely inevitable that cognition sets limits on, say, word length or the total number of phonemes. But such ‘universals’ seem likely to be relatively trivial features of languages, just as may be the case for putative universals in music and other aspects of culture. We should perhaps learn the lesson of Darwinism: a ‘universal’ mechanism of adaptation says little of interest, in itself, about how a particular feature got to be the way it is, or how it works. This truth has dawned on physicists too: universal equations are all very well, but particular solutions are what the world actually consists of, and those particulars are generally the result of contingent history.


Good said...

The paper is in fact a terrible, terrible mess.

At the level of data, it looks like languages have been misclassified: similar languages get different codings, some languages are just misdescribed entirely, and the classification scheme is about 70 years out of date in the first place.

At the level of interpretation, the results have nothing to do with the theories they are supposed to bear on. Literally nothing. The conclusions are non-sequiturs even if we ignore the data problems.

There is extensive discussion on Language Log: http://languagelog.ldc.upenn.edu/nll/?p=3088.

Good said...

By the way, why did you interview only Dryer and Haspelmath, among linguists? If you are writing things like "A generative grammar should show patterns of language change that are independent of the family tree or the pathway tracked through it" (which is not true), shouldn't you have spoken with a linguist who works on generative grammar? Dryer's remarks are sensible, but you have no quote at all from the research community supposedly under primary attack here!

Philip Ball said...

Thanks for these comments. They make a more direct and, to me, comprehensible complaint about the paper than some of those on the Nature blog. I'm perfectly prepared to believe that the Nature editors and referees got it wrong here - that happens. I'm very ready to believe that there are others I could usefully have spoken with, too. I don't pretend to be a linguist, so won't necessarily have found the perfect people in the time available. I guess we all, as reporters, sometimes have to rely on specialists telling us "The person you should really speak to is...". And maybe that didn't happen here. (Surely, however, Dryer is one of those supposedly under attack?)

Good said...

Thank you for your reply. Of course you were correct to speak with Dryer, because his work is being attacked too. But though your article did give voice to Dryer's empirical objections (that the Nature authors didn't examine the right language families), you took at face value the article's other claims: about its implications significance for Universal Grammar and so-called "Chomskyan" linguistics. Yet you appear not to have spoken to anyone who actually pursues these ideas - to see what they think of this reasoning. If you had done so, I think you would have heard incredulity, because (as I wrote in my easlier message) there is no connection at all.

Unfortunately, and sadly, this isn't just a case of ""The person you should really speak to is...". A call to virtually any major university in the UK could have led you to a linguist competent to discuss these matters. In London alone, you could have contacted the linguistics departments at UCL, SOAS or Queen Mary, for example, and asked to speak to any syntax specialist. The problems with the Nature paper are so basic, and the Nature authors' misrepresentation of the opposing position so outlandish, that I doubt you would have found much variation in response. You really could have spoken with almost anyone at all.

I do thank you at least for your openness to discussion here. Much appreciated.

Philip Ball said...

This is curious. I'm not in a position to dispute what you say, except in the one area in which I do have experience: calling a university department at random and asking to speak to a "specialist in X" is pretty much guaranteed to get you nowhere fast. It simply doesn't work that way, believe me.
And I don't want to seem defensive, but when you say that I could have spoken to "almost anyone at all", it is really the case that I have managed to find two of the rare exceptions - one who has written about syntactic universals, the other who has written on Greenbergian word order correlations, yet who are somehow not really qualified to pass judgement here? It may be so - I can't pretend to be familiar with the subdisciplinary subtleties of linguistics. But I do know that in, say, chemistry, someone who has worked on organometallic compounds can be reasonably expected to spot fundamental errors in a paper purporting to be about such compounds. That's to say, organometallic chemists inhabit basically the same world. Again, I'm not disputing; I just find it odd.
That said, it does rather seem as though the Nature review process may have gone badly astray here. Either that, or there is an alarming disagreement in this field about some rather fundamental issues.

Philip Ball said...

By the way, is Steven Pinker just being polite, or is he also uninformed?
Just asking (I know Pinker is the lazy journo's linguist).

Good said...

[part 2 of a 2-part post - having trouble with Blogger.com]

Now the Nature paper comes squarely out of the functionalist "camp". No doubt about that whatsoever. That's the kind of linguist Levinson is, and he has a history of the kind of acrimonious discourse I have discussed above. The most recent case was a 2009 paper on Uniuversals by Levinson with the Australian linguist Nick Evans that had a very similar blizzard of press coverage. It inspired a firestorm of protest from other linguists, with numerous claims (many of them published) that opposing views and even basic facts about particular languages had been misrepresented.

Haspelmath is from the same "camp", and when he discusses generative linguistics, he is often polemical in a manner reminiscent of Levinson. In a better world, no matter what one's views are about scientific questions, you can still be fair to the other side. But still, it seems to me that it's probably a good strategy for a science journalist to seek commentary about the other side of a dispute from...well...from someone who actually is on the other side of the dispute. That emphatically isn't Haspelmath.

This is longer than I'd intended, but I hope it clarifies a bit of the mess that you waded into, no doubt unsuspectingly. And also why it matters to linguists who care about the issues and public perceptions of them.

(As for your question about Pinker, can I just answer that with "no comment"?)

Good said...

[this is part 1]

"[...] it does rather seem as though the Nature review process may have gone badly astray here. Either that, or there is an alarming disagreement in this field about some rather fundamental issues."

I think both explanations are correct.

Roughly speaking, there is an acrimonious divide between linguists who call themselves "functionalists" and those who are called "formalists" (by functionalists) and "generative grammarians" more generally.

In a nutshell: generative grammarians (formalists) believe that fifty years of detailed study of the patterns behind complex linguistic data in the languages of the world has revealed a language faculty -- some of whose pieces appear to be specific to language. Furthermore, the specific structure of this language faculty solves some of the puzzles of language acquisition (discussed in your piece) and restricts cross-language variation (as also mentioned in your piece).

Functionalists are a more diverse group, but most of them probably believe that what is interesting about language is not the formal properties studied by generative grammarians (properties whose very reality is questioned by some functionalists), but the ways in which properties of languages can be attributed to the communicative needs of speakers and the ways in which language variation reflects cultural variation. They believe that *this* is the main message of many decades of linguistic research.

From a rational perspective, there should be no either/or here. Different researchers should be able to find different aspects of language interesting, and debate should be reserved for cases where there are commensurate competing hypotheses about particular facts to sort out. In the actual world of academic linguistics, however, this is not what happens. First of all, an awful lot of the discourse in functional linguistics is devoted to acrimonious complaints about generative linguistics -- often quite misinformed at the most basic level of accurately describing the claims they are attacking (never mind who is right or who is wrong about language itself). Rather than return the favor, most generative linguists tend to simply to ignore the dispute, and go about their work as if the quarrel did not exist at all. Being ignored, as you can imagine, does not make the functionalists any happier with generative linguistics. Result: a split field and a certain failure of productive communication among researchers.

[continued below]

Good said...

Something is wrong with your blog or with blogger.com. It only accepts my posts about 1 out of 30 attempts. Otherwise it either delivers an error message, or else it acts as if the post is about to appear, but it never does.

I tried splitting the post in two, and it ended up accepting part 2 but not part 1. (Is this like the Nature reviewing process...?) That's why you have part 1 after part 2 here.

Philip Ball said...

This is wonderfully helpful - thank you so much. I can now see why, in such a polarized field, what you get depends so much on who you speak to.

It occurred to me that you might have been hoping to find in the news article something more like one might expect in a review process. In general, given the timescales, space constraints and so forth, a news writer generally has to assume that the review process has been conducted properly, and is therefore looking more for a sampling of specialist opinion than for a thorough analysis from the kinds of folks who might have been ideal referees. As it happens, however, I got a bit of both here - the various comments I received were considerably more detailed than the story might have implied, and, in several respects considerably more critical. Had I received them as referees' reports (one virtually was that), I'd definitely not have published the paper. In any event, you're right that I have waded into something unsuspectingly, and have now at least emerged better informed.

Philip Ball said...

And sorry about the problems with blogger.com comments - I don't know what is happening there, but I guess it isn't a problem I can do much about as a humble user of this platform...

Good said...

Happy to help, though I wish the circumstances had been otherwise. Thank you as well for the additional information you offered in your final comment.

By the way, I was serious about London having a wealth of sources for more information, should you ever find yourself needing to write on a similar topic in the future. At UCL, I can recommend Klaus Abels, Ad Neeleman and Andrew Nevins, as well as the "dean of generative linguistics" in the UK, Neil Smith (now emeritus). At SOAS, Peter Sells. And at Queen Mary, David Adger and Daniel Harbour. Further afield within the UK, Caroline Heycock in Edinburgh or Anders Holmberg at Newcastle are also excellent sources. These are some of the people I had in mind when I remarked that "you really could have spoken with almost anyone at all", and I could continue the list with more names too.

Thank you once again.