[Lingtyp] Greenbergian word order universals: confirmed after all
Gerhard Jäger
gerhard.jaeger at uni-tuebingen.de
Mon Nov 6 07:25:13 UTC 2023
P.S. Juergen mentioned the steep learning curve for the phylognetic
comparative method. I am of course always happy to collaborate on
specific projects - just drop me an email.
G.
On 11/6/23 08:23, Gerhard Jäger wrote:
> Dear all,
>
> Thanks to Juegen Bohnemeyer for this thoughtful contribution.
>
> A couple of remarks.
>
> About the phylogenies we used in Jäger & Wahle 2021: We actually did
> not use the world tree. The reason for this choice was the fact that
> branch lengths are an important aspect of the phylogenetic comparative
> method. Branch lengths above the family level in the world tree (the
> one from https://www.nature.com/articles/sdata2018189) are so
> obviously too short that this seriously undermines the method. (Btw,
> the phylogenetic comparative method is fairly robust regarding errors
> in the topology of the phylogenies, but branch lengths are important).
>
> Instead of using the world tree, we inferred a posterior sample of
> phylogenies for each Glottolog family. So like Dunn et al., we used
> separate trees for different families. The main technical innovation -
> besides the fact that we used dozens of families instead of four - was
> that we tested using the same Markov Chain parameters across families,
> where Dunn et al. fitted a separate model for each family. Our's is
> effectively an implementation of Maslova's ideas.
>
> In the extreme case where each family contains just one language in a
> sample, the Jäger & Wahle method is actually equivalent to stratified
> sampling where one language is sampled from each family. So our method
> is not so much an alternative to stratified sampling but an extension
> that allows to use all languages for which you have data.
>
> Regarding the phylogenies derived from ASJP: Juergen, you say,
> "Notably, Jäger’s method takes into account systematic sound
> correspondences, so to my uninitiated eye, it looks like an
> unassailable computational implementation of the historic-comparative
> method." I am afraid you are doing me too much honor here. My method
> takes recurring sound correspondences into account, but it is still a
> far cry from a full implementation of the historic-comparative method.
> However, in Jäger 2018 (https://www.nature.com/articles/sdata2018189),
> I ran a lot of tests demonstrating that the quality of the phylogenies
> obtained this way is fairly adequate (using Glottolog as gold
> standard) despite the method's conceptual shortcomings.
>
> I agree with Juergen that there is no consensus approach for the
> phylogenetic comparative method. I am not sure though that such a
> consensus is possible or even desirable. A linguistic result - say,
> the universal correlation between adposition-noun order and
> verb-object order - is trustworthy if several different methods
> confirm it. These methods may include stratified sampling with
> different sampling protocols, family bias estimation, phylogenetic
> comparative methods with different trees and different priors, perhaps
> others. If some approaches confirm it and others do not, that is a
> reason to dig deeper.
>
> Best, Gerhard
>
>
>
> On 11/6/23 00:22, Juergen Bohnemeyer wrote:
>>
>> Dear all – Two points:
>>
>> First, this thread has served to alleviate somewhat my concerns about
>> the review processes used by the generic science journals – mostly
>> through offline one-on-one communications I’ve received. Thanks!
>>
>> Secondly, I’d like to, maybe gently push back, or may be just
>> elaborate, on the contrast Simon Greenhill draws between phylogenetic
>> inference and the “overly simplistic correlation methods”, as he
>> calls them. In the process, I’d like to try to clarify what in my
>> mind are the strengths and weaknesses of each (type of) method. And
>> I’ll pose a couple of challenge questions to the detractors of
>> phylogenetic/coevolutionary testing and one for their proponents, and
>> I hope all of these questions won’t go unanswered 😊
>>
>> Before I do that, let me flag my interest in this: I’m currently in
>> Leipzig spending a sabbatical with Russell Gray’s Department of
>> Linguistic and Cultural Evolution (≈ the Grambank team), with the
>> principal objective of figuring out how to capitalize the most on
>> Grambank for my project. The project tracks the typological
>> distribution of functional expressions and attempts to account for it
>> (the distribution) in terms of an evolutionary process that makes
>> grammaticalization sensitive to functional pressures. Since coming
>> here, I’ve been struggling with the question whether I should adopt
>> stratified sampling, a phylogenetic method (I’m not going to use the
>> term ‘co-evolution’ here, because I’m at this stage of the project
>> just looking at the emergence of individual functional expressions),
>> or both. So I’m airing my thoughts very much in hopes of eliciting
>> feedback.
>>
>> So. Let’s start with a very basic reminder: We cannot observe
>> causality, we can only infer it, as David Hume taught us (i.e.,
>> humanity) nearly three centuries ago. The difference between the
>> methods at issue is one between synchronic inference and diachronic
>> inference.
>>
>> Now, before I go any further, a big disclaimer: I think I understand
>> at the most abstract level the ideas underlying phylogenetic
>> inference. But I don’t understand the statistical “sausage making”
>> involved. Meaning I have many basic questions about the algorithms
>> used for phylogenetic inference.
>>
>> With that said: Imagine an ideal world in which we had perfect and
>> reliable knowledge about the complete phylogeny of every extant human
>> language and every language that ever existed. It seems indisputable
>> to me that in this world, we would test hypotheses about causal
>> relations between two linguistic features not merely by looking at
>> their synchronic distributions, but also by looking at whether the
>> two features tend to co-evolve across phylogenies or whether they
>> appear to have merely accidentally “travelled together” through time
>> in a few families.
>>
>> So this would be my first question to the detractors of
>> phylogenetic/dynamic (Maslova’s term)/co-evolutionary tests for
>> typological generalizations: do you disagree that if we had that
>> perfect knowledge, we would as a matter of course take into account
>> the phylogenetic perspective?
>>
>> Now, in the real world, our knowledge of phylogenies is itself
>> largely inferred from synchronic distributions. Crucially, though,
>> these inferences are based on data that may overlap with the
>> typological patterns of interest, but is largely independent of it.
>> To make this more concrete:
>>
>> * Jäger & Wahle (2021) use the world tree inferred from ASJP
>> <https://asjp.clld.org/> cognate data presented in Jäger (2018).
>> Notably, Jäger’s method takes into account systematic sound
>> correspondences, so to my uninitiated eye, it looks like an
>> unassailable computational implementation of the
>> historic-comparative method. Of course, the data the analysis is
>> based on is another matter.
>>
>> * Verkerk et al. (ms.), which Martin cites in the post that started
>> the thread, rely on Glottolog, which Jäger & Wahle also used to
>> validate their inferred world tree. In my understanding,
>> Glottolog is sourced by a compilation of the best available
>> evidence for phylogenetic relations, principally curated by its
>> lead author, Harald Hammarström.
>>
>> Two more points to keep in mind here:
>>
>> * Stratified sampling of course also takes into account
>> phylogenetic and areal information. So the two types of
>> approaches have really more in common than Simon gives them
>> credit for in his reply, I think. However, stratified sampling
>> uses phylogenetic and areal information in first approximation
>> only “negatively”, as it were: by systematically removing
>> observations from the analysis that are suspected to be “tainted”
>> by areal or genealogical dependencies.
>>
>> * The distance of our real world from that ideal world in which we
>> have perfect knowledge of phylogenetic and areal relations is in
>> my (again, uninitiated, in the sense that I’m not a historical
>> linguist) mind pretty darn great. However, the phylogenetic
>> statistics algorithms use “forests” of alternative phylogenetic
>> trees weighted for the confidence the field (or the analysts)
>> has/have in them, factored into the tests as Bayesian priors. In
>> this sense, the use of phylogenetic information is more
>> sophisticated in the dynamic tests than in anything based on
>> stratified sampling that I have seen.
>>
>> So with all this assembled, here’s my semi-informed take on the pros
>> and cons of each approach:
>>
>> * Stratified sampling (I’m going to start with the method that has
>> been practiced the longest)
>>
>> o Pro
>> + Technologically and conceptually simple, easy to
>> implement by anybody with the most basic level of
>> training in inferential statistics, and transparent for
>> the analyst to track and understand the effects of their
>> decisions.
>> + Statistically conservative, in the specific sense that it
>> minimizes the odds of false positives, i.e., of support
>> for generalizations for which the evidence isn’t really
>> there. (One might think that Dunn et al. 2011 present
>> counterevidence to this. But the apparent false positives
>> pointed out by Dunn et al. turn out to be true positives
>> after all, now that Jäger & Wahle and Verkerk et al. have
>> replicated their analysis with a larger sample.)
>> + Validity of testing depends only on one key assumption –
>> the big one: independence of observations. (Detractors of
>> stratified sampling might however argue that it is
>> impossible to ensure that this assumption is warranted in
>> any decent-sized sample.)
>> o Con
>> + There has never been a consensus on an optimal sampling
>> algorithm that balances minimization of areal and
>> genealogical biases with retaining enough statistical
>> power to detect the patterns of interest. With such a
>> consensus absent after 45 years of some of the smartest
>> minds in the field applying themselves to the problem,
>> maybe there is no optimal solution to be had?
>> + Based on information reduction (exclusion of
>> observations); does not take into account phylogenetic
>> information to the fullest measure of what can be
>> concluded from it.
>>
>> * Dynamic/phylogenetic/co-evolutionary inference
>>
>> o Pro
>> + Utilizes the best available phylogenetic information “for
>> all it’s worth”, directly examining whether features tend
>> to coevolve across lineages or only in particular
>> lineages. (The same cannot currently be said about areal
>> information. However, Verkerk et al. (ms.) use mixed
>> models regression – which is feasible based on Grambank
>> thanks to its relatively low percentage of missing
>> observations – to factor in possible areal effects.)
>> + Does not depend on the assumption of independence of
>> observation.
>>
>> o Con
>> + Very steep training demands both on the conceptual and on
>> the technological side. Currently only a handful of
>> specialists are able to perform such analyses. In
>> addition, the computational complexities are so demanding
>> as to render the analyses effectively opaque for all but
>> those specialists, raising questions about the validation
>> of such tests. (Gerhard Jäger’s comparison to carbon
>> dating in archeology is apt here as well, I think: while
>> phylogenetic analysis will likely become more standard in
>> the future, it may be the case that typologists will
>> always have to rely on specialists for actually
>> performing such tests, just like archeologists rely on
>> physicists or lab technicians for performing carbon
>> dating. I think 😉)
>> + Reliability of phylogenetic inference seems to depend on
>> the size of the phylogenies involved: the smaller the
>> family, the higher the uncertainty in the absence of
>> actual historic data.
>> + As with stratified sampling, there does not yet appear to
>> be a consensus approach. Jäger & Wahle (2021) and Verkerk
>> et al. (ms.) use different algorithms (implemented in
>> different software packages), make different assumptions,
>> and their tests are conceptually quite distinct.
>> + Outcomes depend on many assumptions which are surrounded
>> by varying degrees of uncertainty.
>>
>> So at long last, here are my three challenge questions:
>>
>> * For the detractors of phylogenetic inference:
>> o (Repeated from above to make sure this doesn’t get lost) Do
>> you (dis)agree that if we had complete and robust knowledge
>> of the phylogenies of the extant languages, we would perform
>> phylogenetic tests as a matter of course when attempting to
>> validate typological generalizations?
>> o Do you dispute that since phylogenetic testing of typological
>> generalizations has become possible, typology as a field
>> cannot simply ignore it and move on without it?
>> * For the proponents of phylogenetic inference:
>> o Do you dispute that stratified sampling is statistically
>> conservative and has the advantage of transparency and
>> robustness vis-à-vis the underlying assumptions, so it is
>> likely not a method that typology will simply move beyond and
>> disregard in the future?
>>
>> And there it. Apologies for the overlong post! – Cheers – Juergen
>>
>> Juergen Bohnemeyer (He/Him)
>> Professor, Department of Linguistics
>> University at Buffalo
>>
>> Office: 642 Baldy Hall, UB North Campus
>> Mailing address: 609 Baldy Hall, Buffalo, NY 14260
>> Phone: (716) 645 0127
>> Fax: (716) 645 3825
>> Email: jb77 at buffalo.edu <mailto:jb77 at buffalo.edu>
>> Web: http://www.acsu.buffalo.edu/~jb77/
>> <http://www.acsu.buffalo.edu/~jb77/>
>>
>> Office hours Tu/Th 3:30-4:30pm in 642 Baldy or via Zoom (Meeting ID
>> 585 520 2411; Passcode Hoorheh)
>>
>> There’s A Crack In Everything - That’s How The Light Gets In
>> (Leonard Cohen)
>>
>> --
>>
>> *From: *Lingtyp <lingtyp-bounces at listserv.linguistlist.org> on behalf
>> of Simon Greenhill <simon at simon.net.nz>
>> *Date: *Saturday, November 4, 2023 at 13:39
>> *To: *lingtyp at listserv.linguistlist.org
>> <lingtyp at listserv.linguistlist.org>
>> *Cc: *Dunn Michael <michael.dunn at lingfil.uu.se>, Russell Gray
>> <russell_gray at eva.mpg.de>
>> *Subject: *[Lingtyp] Greenbergian word order universals: confirmed
>> after all
>>
>> Colleagues, Martin, everyone else
>>
>> Thank you for sharing your perspectives on our 2011 paper. It's nice
>> to see this still be discussed more than a decade later. However, I
>> would like to express my concerns and disagreements with some of the
>> points you've raised.
>>
>> I'm very proud of the Dunn et al. paper for a number of reasons. I'll
>> name three.
>>
>> First, the paper showed that the overly simplistic correlation
>> methods that had been used to make sweeping global claims were
>> problematic. We need better tools to tackle these questions, and the
>> tools we applied were one part of a better toolkit.
>>
>> Second, it highlighted the need to understand language systems in a
>> diachronic manner. We cannot decouple language typology from language
>> history, instead we need to understand how these are entangled.
>>
>> Third, it emphasised the way that particular configurations of
>> languages can be arrived at via different routes in different
>> families at different times. This enables a much richer understanding
>> of how these particular generalisations have arisen.
>>
>> Have Jäger and Wähle disproved any of that? no. Maybe these were not
>> completely novel insights (Maslova’s work has been mentioned which
>> touches on a few of these issues too, for example), but these ideas
>> did appear to crystallise in this paper.
>>
>> While it's certainly important to revisit and reevaluate research
>> findings to ensure accuracy, it is crucial to approach these
>> discussions with an understanding of the scientific process.
>> Scientific paradigms evolve over time, and different studies may
>> yield varying results due to changes in methodologies, data sources,
>> and sample sizes. This doesn't necessarily imply that the initial
>> research was flawed or that the authors were neglectful. In
>> particular, the tools, the data, and our understanding of how
>> languages change are substantially further advanced than they were a
>> decade ago (or, I know that *my* understanding of these things is
>> more advanced now, at least). And these other papers that you mention
>> -- and many other studies -- have built upon the work we did in 2011.
>>
>> Furthermore, I would like to caution against drawing overly broad
>> conclusions about the quality of research published in high-prestige
>> journals. The peer-review process in such journals is rigorous, and
>> while they may occasionally feature sensationalist claims, this
>> doesn't diminish the overall value they contribute to the scientific
>> community. For the record, of the handful of papers I've had in these
>> journals *all* have been reviewed by people I would infer to be
>> linguists based on the comments and issues they raised. We did not
>> send these papers to these journals to avoid linguistic reviewers
>> but, frankly, I've had better reviews at these journals than at
>> prominent linguistics journals (and by "better" I mean more rigorous,
>> more thorough, and more critical).
>>
>> Finally, linguistic typology is an ongoing and evolving field trying
>> to tackle very difficult problems. We need all the tools and
>> approaches we can get to solve these problems across all the levels
>> that languages operate on (from detailed language internal analyses
>> to high-level global analyses). Rather than looking back and
>> gate-keeping what is 'real’ typology published in ‘real’ linguistics
>> journals, we should shift our focus forward. Typology can be a
>> welcoming and diverse community that embraces a wide range of
>> approaches, analyses, and styles. Let's look outward to foster
>> connections with other fields and disciplines.
>>
>> After all, why shouldn't linguistic typology work be everywhere in
>> science? it's certainly interesting enough.
>>
>> Simon
>>
>> Dr. Simon J. Greenhill
>>
>> Associate Professor
>>
>> Te Kura Mātauranga Koiora | School of Biological Sciences
>> Te Whare Wānanga o Tāmaki Makaurau | University of Auckland
>>
>> Abteilung für Sprach- und Kulturevolution | Department of Linguistic
>> and Cultural Evolution
>> Max-Planck-Institut für Evolutionäre Anthropologie | Max Planck
>> Institute for Evolutionary Anthropology
>>
>> _______________________________________________
>> Lingtyp mailing list
>> Lingtyp at listserv.linguistlist.org
>> https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Flistserv.linguistlist.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Flingtyp&data=05%7C01%7Cjb77%40buffalo.edu%7Cbe1181a9879e434ccc0308dbdd3319c4%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638346983806627973%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=OgjDO%2FGow%2FfFmxOIvmrWNNq1nwcsZaLzI5OMnoGjBXg%3D&reserved=0
>> <https://listserv.linguistlist.org/cgi-bin/mailman/listinfo/lingtyp>
>>
>>
>> _______________________________________________
>> Lingtyp mailing list
>> Lingtyp at listserv.linguistlist.org
>> https://listserv.linguistlist.org/cgi-bin/mailman/listinfo/lingtyp
>
> --
> Prof. Dr. Gerhard Jäger
> Universität Tübingen
> Seminar für Sprachwissenschaft
> Tel.: +49-7071-29-77302
> http://www.sfs.uni-tuebingen.de/~gjaeger/
--
Prof. Dr. Gerhard Jäger
Universität Tübingen
Seminar für Sprachwissenschaft
Tel.: +49-7071-29-77302
http://www.sfs.uni-tuebingen.de/~gjaeger/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.linguistlist.org/pipermail/lingtyp/attachments/20231106/52e421d0/attachment-0001.htm>
More information about the Lingtyp
mailing list