<!doctype html public "-//W3C//DTD W3 HTML//EN">
<html><head><style type="text/css"><!--
blockquote, dl, ul, ol, li { padding-top: 0 ; padding-bottom: 0 }
--></style><title>Re: Dunn et al. on word order typology in
"Nature"</title></head><body>
<div><font face="Times New Roman" color="#000000">We should not lose
sight of what is good in Dunn et al.'s Nature article. Many
typologists, starting with Greenberg, have argued that synchronic
language universals are really just manifestations of diachronic
universals. The state-process model used by Dunn et al. (see p. 5 of
their supplementary materials) has been used by Greenberg (1978) and
Maslova (2000). Dunn et al.'s method (and also Maslova's) allows us to
take a major problem with sampling - historical dependence - and
exploit it to uncover valid language universals. Using quantitative
and statistical techniques will allow us to make more precise
generalizations and assign a degree of goodness of fit to our
theoretical models of typological universals.</font></div>
<div><font face="Times New Roman" color="#000000"><br></font></div>
<div><font face="Times New Roman" color="#000000">Nevertheless, like
others here, I am unconvinced of the results due to problems with the
way they apply the method. Dunn et al.'s analysis (pp. 5-6 of the
supplementary materials) treats two models as mutually exclusive: an
"independent" model, in which a word order switches
independently of other orders, possibly with a weighted preference for
one order; and a "dependent" model, in which two word orders
are linked and change together. These competing models are tested
against each phylogeny (family tree), with one or the other winning
out (*very* crudely, if branches with linked changes outnumber
branches with single order changes, then the test takes the word-order
correlation as justified for the tree). But this does not test the
model that most typologists assume; at best it tests the
oversimplified model of Theo Vennemann and Winfrid Lehmann from the
1970s, which reduced the diversity of word order in the world's
languages to just two types, VO and OV. (This is also basically the
generative head-ordering parameter.) But from Greenberg's original
paper (Greenberg 1966) onwards, most typologists have adopted a model
in which both single-order preferences (Greenberg's 'dominance') and
linkages (Greenberg's 'harmony') compete with each other and jointly
determine patterns of word order variation, as Matthew noted in his
last post. This can be demonstrated statistically: Justeson and
Stephens (1990) did a log-linear analysis on a large synchronic
language sample and showed that the best-fit model included both
dominance and harmony factors. I presume there is a dynamic equivalent
of log-linear analysis that could be used to test the model that most
typologists accept.</font></div>
<div><font face="Times New Roman" color="#000000"><br></font></div>
<div><font face="Times New Roman" color="#000000">Regarding the issue
of lineage-specific vs. universal patterns: Keith Poole and I had to
deal with a similar question in using multidimensional scaling to find
universals of grammatical categories (Croft and Poole 2008). There,
the contrast was between language-specific MDS models (as used, for
example, by Barbara Malt and her colleagues) and crosslinguistic MDS
models. We argued that if the regularities were culture-specific, then
mixing in languages with culturally-specific category structures would
reduce the goodness of fit of the MDS model; but if the regularities
were crosslinguistic (i.e. universal), then mixing together languages
would improve the goodness of fit of the model. (We found the latter.)
The same presumably would apply for word order patterns, mutatis
mutandis. But here the problem with the Dunn et al. result is that
they test only four lineages, representing only 7.5% of language
genera (low-level language families) in the world; the other 92.5% of
language genera occur in other lineages. Also, for many of the
pairwise correlations they test, including the two illustrated in
their article, Bantu is too shallow a family to exhibit any variation;
and for the two illustrated in their article, even in the other
families very few branches undergo a word order shift. So the
empirical sample, though well distributed geographically, is very
small and has few independent changes to evaluate. And here we hit a
problem that all researchers on language universals hit: our
uncertainty about language phylogeny in most parts of the world,
especially for deeper families, but even for subgrouping in accepted
families. In this latter area, quantitative methods are also being
applied; but that is another story.</font></div>
<div><font face="Times New Roman" color="#000000"><br></font></div>
<div><font face="Times New Roman" color="#000000">I think typologists
should welcome this effort to marry phylogeny and typology, even if we
remain unconvinced of the particular result in this
paper.</font></div>
<div><font face="Times New Roman" color="#000000"><br>
Bill<br>
<br>
<br>
Croft, William and Keith T. Poole. 2008. Inferring universals from
grammatical variation: multidimensional scaling for typological
analysis.<i> Theoretical Linguistics</i> 34.1-37.<br>
<br>
</font><font face="Times" color="#000000">Greenberg, Joseph H. 1966.
Some universals of grammar with particular reference to the order of
meaningful elements.<i> Universals of Grammar</i>, ed. Joseph H.
Greenberg, 2nd edition, 73-113. Cambridge, Mass: MIT Press<br>
<br>
Greenberg, Joseph H. 1978. Diachrony, synchrony and language
universals.<i> Universals of Human Language, Vol. 1: Method and
Theory</i>, ed. Joseph H. Greenberg, Charles A. Ferguson and Edith A.
Moravcsik, 61-92. Stanford: Stanford University Press.<br>
<br>
Maslova, Elena. 2000. A dynamic approach to the verification of
distributional universals.<i> Linguistic Typology</i>
4.307-33.</font><br>
<font face="Times" color="#000000"></font></div>
</body>
</html>