Dunn et al. on word order typology in "Nature"

[Søren Wichmann]

The model of coevolution was developed in biology, where it is highly 
meaningful to test whether features that seem to be correlated change in 
tandem along phylogenetic lineages, since biological traits normally 
don't spread laterally (bacteria providing a major exception). But for 
linguistic typological features we know that there is a strong areal 
effect, so if this effect is not somehow built into the evolutionary 
model one can only expect results that are difficult to interpret. It is 
hard to expect correct conclusions from the wrong premises. If the 
premise is that correlated linguistic feature should coevolve 
historically and the conclusion is that this evolution is 
lineage-specific then we should pause before accepting this conclusion. 
If we are to adopt methods from biology when studying linguistic 
typology we should be careful about which models to adopt.

Søren.

Michael Dunn wrote:
> Thanks for your comments, Matthew.  And thanks Jess Tauber for putting
> up a link. Here's another link that should give free access to both the
> paper and the supplementary materials (also important, given the brevity
> of Nature papers): http://www.mpi.nl/publications/escidoc-95245/
> 
> I'd like to respond to Matthew's comments on the list, because I think
> they touch on some issues which are important to all linguistic
> typologists.
> 
> The claim that some word order dependencies should only exist ACROSS
> families but not WITHIN them needs unpacking. Unrelated languages are
> products of historical processes too, it's just that these historical
> processes are unknown or (if you use sampling methods) ignored. The
> Phylogenetic Comparative Method test for dependency used in our paper
> (described in the Supplementary Materials) controls for genealogical
> relatedness by looking within a lineage and inferring the extent to
> which changes in two features are 'coupled'. In some cases, changes in
> one feature are regularly associated with changes in another, in other
> cases they're not. For the well-known correlation between verb-object
> order and Adposition order, phylogenetic comparative methods do detect
> the correlation within the Indo-European and Austronesian families. But
> the correlation is absent in Uto-Aztecan, despite that fact that
> verb-object orders and adposition orders do change within the history of
> the family.
> 
> It could be argued I suppose that this is a statistical universal. But
> the usual understanding of statistical universal (e.g. Dryer 1998) is
> that statistical universals are universal tendencies: the tendency
> itself should be present universally. If the tendency is absent in some
> lineage then it's not a statistical universal (unless you wish to admit
> statistical statistical universals!), it's a lineage specific process.
> 
> As to the unexpected correlations, Dryer 2007 ("Word Order", in the
> Shopen trilogy) makes strong statements about what word order features
> show intercorrelation and what don't. Our results (Figure 2) show that
> this differs from lineage to lineage. A correlation between SV and OV
> exists in Uto-Aztecan, but not any of the other families. There is no
> correlation between order of adjective and noun and relative clause
> order in Indo-European (although there is in Austronesian and
> Uto-Aztecan), but Indo-European alone does have an evolutionary
> correlation between adjective-noun and genitive-noun orders.
> 
> So far I've just been talking about tests for the existence of
> dependencies between particular pairs of features. An additional thing
> we get for free from the comparative phylogenetic approach is an
> explicit model of evolutionary change for each feature. Where a pair of
> features are correlated, the method infers which changes between states
> are more or less probable.  Figure 3 in the paper shows the different
> models of transition probabilities inferred between VO and adposition
> states in Austronesian and Indo-European. The patterns of evolutionary
> change inferred in the two families are different from one another:
> these features are dependent in both families, but it's not the same
> dependency. We can do this for any grammatical dependency detected in a
> family, something I am quite excited about for future work.
> 
> I hope this somewhat clarifies things for LingTyp readers. I would be
> happy to expand on this discussion if there's interest.
> 
> Best, Michael
> 
> 
> On Wed, Apr 13, 2011 at 06:59:39PM -0400, Matthew Dryer wrote:
>> I am not sure whether this list is the appropriate venue for
>> commenting on the paper in Nature by Michael Dunn et al.  But since
>> it is, as Martin said, unusual for a typological paper to appear in
>> Nature, and since the paper was brought up on this list, I think
>> some very brief comments are in order.  My apologies to people on
>> the list who have not had an opportunity to read the paper.
>>
>> Put briefly, the paper is based on major misunderstandings of the
>> claims of word order typology.  All of the results of the paper are
>> already familiar to me and are entirely consistent with claims that
>> have been made in the word order literature.  If I can take the
>> liberty of quoting Michael's own words from his email,
>>
>> "(i) don't find many of the expected correlations"
>>
>> What the paper shows is that we often don't find the expected
>> correlations WITHIN language families.  But there are many reasons
>> why we should expect this, and nothing in the word order literature
>> would lead us to expect otherwise.  We only expect to find the
>> expected correlations ACROSS families.
>>
>> "(ii) find many correlations which were unexpected"
>>
>> In fact, the correlations of this sort mentioned in the paper are
>> well-known, such as the correlation between the order of adjective
>> and noun and the order of relative clause and noun.  There is
>> nothing unexpected about these correlations.
>>
>> "(iii) find that even where dependencies are found between the same
>> pairs of features in two lineages, the evolutionary models
>> underlying these dependencies are different"
>>
>> There are potentially novel results here, but I see no reason to
>> think that these differences are due to anything other than random
>> variation.
>>
>> Michael Dunn wrote:
>>> Thanks Martin for the kind words!
>>>
>>> We've put together some materials for non-experts to help with
>>> understanding the paper: http://language.psy.auckland.ac.nz/wordorder/
>>> I'm sorry that these are not currently at the appropriate level for
>>> typologists, but we'll add to them as necessary.
>>>
>>> I don't quite get your parenthesis at the end: one of the points of the
>>> paper is that we use a particularly stringent and statistically powerful
>>> control for genealogical relatedness, and nevertheless (i) don't find
>>> many of the expected correlations, (ii) find many correlation which were
>>> unexpected, and (iii) find that even where dependencies are found
>>> between the same pairs of features in two lineages, the evolutionary
>>> models underlying these dependencies are different.
>>>
>>> Best, Michael
>>>
>>> On Wed, Apr 13, 2011 at 10:06:41PM +0200, Martin Haspelmath wrote:
>>>> In my recollection, this is the first typology article to be
>>>> published in Nature:
>>>> http://www.nature.com/nature/journal/vaop/ncurrent/full/nature09923.html.
>>>> Congratulations to our colleagues in Auckland and Nijmegen!
>>>>
>>>> There is also a popularized account in Nature News
>>>> (http://www.nature.com/news/2011/110413/full/news.2011.231.html),
>>>> and a Nature editorial about "Universal truths"
>>>> (http://www.nature.com/nature/journal/v472/n7342/full/472136a.html).
>>>> How wonderful to see that typology has become so important!
>>>>
>>>> (Unfortunately, I don't see what is new in the paper -- maybe
>>>> someone can explain this? Didn't we know all along that we are not
>>>> likely to get correlations if we don't control for genealogical
>>>> relatedness?)
>>>>
>>>> Martin
>