Talk:Mass comparison/Archive 1

Meaning of "mass", and onomatopoeias
Hi Mustafaa,

Thanks for creating the separate page. I have added my understanding of the "mass" in "mass comparison"; is it accurate enough?

I have put back the English onomatopoeia examples only to illustrate the concept for the benefit of non-linguist. Do you object to them because they are bogus, or because they seemed to be presented as examples of the problem? Is the present wording OK?

(The fact that English is particularly rich in onomatopoeic words is not questioned, is it? Even the most nationalistic Romance-speaking cartoonists often feel compelled to use English words for their "sound effects".  If that is not "hard scientific proof", I don't know what is... 8-)

BTW, the paragraph that summarizes the traditional method has become too heavy and waffling with all those examples and non-examples. Would you mind if I trim it a bit? The point was only to illustrate the use of the classical method. (But now I am curious -- how were the Sino-Tibetan and Afro-Asiatic families established?)

All the best, Jorge Stolfi 20:54, 4 May 2004 (UTC)

The "mass" part looks fine, and I agree that the English examples are useful; it was only the phrasing that I thought was misleading (insofar as it suggested that these were among the "basic terms" he would have been looking through.)

Richness in onomatopeic words is relative; English doesn't hold a candle to Japanese, for instance, and I don't think it's got any more than Maghreb Arabic... I don't honestly know whether it's richer or less rich than average. But I agree, that sounds like excellent proof that it's richer than Romance!

Sino-Tibetan was, in strictly orthodox terms, not "established" until the etymological dictionaries of Starostin (and Matisoff?) last decade; but in reality, it became universally accepted in the forties, as a result of Shafer and Benedict's work. Afro-Asiatic now has two etymological dictionaries (with mutually contradictory sound shifts listed) - Orel & Stolbova, and Chris Ehret; but it was accepted well before that, because the compelling morphological and pronominal similarities between the families (other than Omotic), as well as some shared roots, are so obvious that even medieval scholars commented on some of them. - Mustafaa 21:29, 4 May 2004 (UTC)

There was an invisible comment
 * milk? wind???  How are these remotely onomatopoeic? I suppose if one says them with a suitably gurgly or breathy voice, respectively?

I am not sure about "wind", but the suggestion that "milk" may be onompatopoeic comes from a Scientific American article on world languages (possibly by Greenberg himself) of many years ago. The idea was that it may have come from the gurgling sounds made by suckling babies. IIRC, the article mentioned words in other language families with similar sounds but meanings related to throat, swallowing, and things like that. Jorge Stolfi 02:14, 16 Sep 2004 (UTC)


 * I guess that "wind" could theoretically be onomatopoetic in that it imitates a blowing sound. "wiieen" or something.

Yeh, well clearly if they ARE onomatopoetic in origin, we are not talking about the sound of the modern word, but of the PIE root or perhaps something even further back. --Doric Loon 21:30, 28 March 2006 (UTC)


 * >The fact that English is particularly rich in onomatopoeic words is not questioned, is it?< Sorry, but I think this is rubbish. Garik 20:26, 15 May 2006 (BST)

Mis-conceptualisation of language descent
The idea of language descent as something that happens smoothly and gradually is flawed. The text about relations between languages being invisable after 10000 years of divergence demonstrates the problem…


 * If speakers of language A and speakers of language B come into contact, and within the next generation or two, speakers of A and B become bilingual, and a couple generations later language B becomes the dominant language for whatever social reason (even some benign social reason, such as the two groups considering themselves one group and group B represents the majority of the people in question), you have this picture (-> represents a generation gap or two):
 * language A speakers -> speakers of A & B (but with some influence from A) -> speakers of languages A & B, with mutual influence -> speakers of language B, with lots of influnce from language A / speakers of language C
 * language B speakers -> speakers of B & A (but with some influence from B) -> speakers of languages B & A, with mutual influence -> speakers of language B, with lots of influence from language A / speakers of language C

This is a relatively balanced scenario, and of course actual situations are much less balanced.

So the resulting language C could be said to be genetically B, with lots of influence from A, but at the same time, it could be said to share features from both languages, and be genetically both (probably with some majority of genetic relation going back to B).

At this point, saying that the resulting language C is "descended" from either A or B isn't inaccurate, so why call it genetically n/either? In this case, many of the features from both originating languages is obliterated rather quickly. Since this sort of thing happens all the time (speakers of different languages coming in contact with each other), all evidence but a few words for the descent from one language could quickly be wiped out. Not only is my argument meant to show that any estimate should be well under 10000 years, but that no language is completely free of contact with other languages and so estimates and theories of language change shouldn't be based on that.


 * Hi there, Firespeak. You are quite right that it can happen very fast.  It can also happen very slowly.  Icelandic is a language which has barely changed in the last 1000 years.  Most linguists today would say that this is so variable that you can't put a figure to it.  --Doric Loon 22:17, 5 August 2005 (UTC)
 * Having thought a little more about this, I decided to delete that sentence. It wasn't helping the argument and can as easily be omitted.  The alternative would be to discuss how controversial it is, but since it is not the main point here, that would be a diversion from the purpose of the paragraph.  The theory had a wider applicability, in the eyes of those who espoused it.  They thought they could date hypothetical parent languages from the degree of divergence between the daughter languages.  I think no-one today still believes this. --Doric Loon 22:25, 5 August 2005 (UTC)
 * Indeed, describing the evolution of languages by a tree is a gross simplification. The tree model was an (unconscious?) ripoff by linguists from Darwin's theory for biological species. It sort of works in biology --- only because species, by definition, cannot interbreed.  (But even there it isn't quite correct --- bacteria can swap genes promiscuously, and hybrids of higher species are sometimes viable.) But languages do mix: it happens all the time with lexicons, but sometimes also with phonetics and grammar --- isn't that so? However, this creates a problem for the comparative method as well as for mass comparison. Jorge Stolfi 17:15, 17 December 2005 (UTC)


 * I put back the sentence about the comparative method being limited to 10.000 years, with different wording. It should now be clear that it is not a theoretical estimate, but merely an observation of the current state of the art. Correct me if I am wrong, but all the widely accepted language families that have been found by this method are believed to have split from a common ancestor within that time span. (the proto-Nostratic may be older, but it has not been widely accepted, right?) Note that sayng "random changes make the method unusable beyond 10,000 years" is very different from saying "random changes accumulate at a constant rate" (which, agreed, is definitely false for many  historically attested languages). Jorge Stolfi 20:14, 17 December 2005 (UTC)

I wonder if it might be possible to calibrate Greenberg's technique differently depending on those exogenous factors that can be reasonably assumed to influence the descent of languages. For example, MLC has been calibrated on the Indo-European family, surely the most studied in all of comparative linguistics, and found to have excellent agreement with the more traditional results. But in the Indo-European family, there was (and still is) extensive - indeed, overwhelming - contact between speakers of different languages; whether due to warfare or trade, these contacts have involved thousands of individuals over short time spans. The miscegeny of language under these conditions is all but unavoidable, even between groups of speakers who would rather die than admit any sort of kinship to the other. In this instance, geography and population dynamics has played a significant role in propelling many languages into mutual proximity. Moreover, the availability of lexicons is phenomenal, even for languages that have ceased to be used (e.g., the Gothic languages, or the extinct Celtic tongues).

Now consider his use of MLC to categorize the African tongues. Here the traditional comparative method had to "catch up" to MLC, but in the end, the categorizations of MLC have been, for the most part, ratified by the more traditional comparative approach. In this instance, the lexicon availability is not what it was for the IE test, but due to extensive colonization and "human trade" (a/k/a slavery) at the hands of numerous European communities, information was precipitated from oral to written form broadly and over a period of several hundred years, before some of these languages ceased to exist. The mixing factors that promote contact and exchange in this scenario are still trade and warfare - but now you tend to have smaller groups of folks interacting at any given time, and over a larger geographic range.

Now we turn to the Americas. Here the method stands the least chance of convenient calibration, because geography and sheer smallness of populations conspire against it. Moreover, it might be that the availability of reliable lexicons is weakest for precisely these languages (all 1500 of them, give or take). Look at it this way: A finely-tuned Rolls-Royce engine will still perform poorly if it is fueled by low-grade fuel in which a significant water fraction exists. Does it mean the engine is sub-standard, or that suspicion and mistrust should be attached to its performance when it is properly fueled? By no means!

In addition to the obvious notion of viewing language evolution in a genetic sense (including gene mixing and "splicing"), it might help to think of languages in large or highly compartmentalized geographic areas as evolving along some sort of "mean free path" axis. This is certainly true of gases mixing in a vessel - at any level of partial pressures, the tendency of gases to react and exchange electrons and re-align orbital configurations is going to be a statistical function of their tendency to "collide." In other words, we might be able to impute an almost "kinetic" component to language evolution, and perhaps even thermodynamics as well (what are the barriers to linguistic exchange and/or self-modification?).

Referring back to my observation two paragraphs ago: Do we throw out the MLC baby with the "dirty lexicons" bathwater? Maybe a variant of MLC can be calibrated against languages like some of the trans-Caucasian tongues, or some other groups where mutual interactions have been infrequent over time. And certainly more accurate lexicons cannot help but improve the quality of the phylogeny imputed by MLC for the American languages.

In the spirit of intellectual honesty and fairness, I must disclose that I am both a Stanford graduate *and* a big fan of Joseph Greenberg (though I never had the privilege of meeting the man). That said, I hope the preceding observations and proposals may be of some interest and assistance in resolving this controversy. I am also *not* a linguist by training or profession, so I apologize if anything (or everything) I've said is merely re-stating the obvious.

Thanks! I feel much better now!

Jorge, it's good to see you active in this forum. You may recall that I was part of the Voynich discussion group for several months a few years ago.

Kevin K Gillette 20:59, 28 March 2006 (UTC)

Reconstructed Proto-Franco-Sino-Indonesian
This linked article is funny, but I don't think it portrays mass comparison, and in fact it doesn't mention the term.

Mass comparison means comparing a large number of languages, and is meant to find which of these languages are more closely related to each other than to the other languages in the comparison. The humorous example compares only 3 languages and doesn't comment on which 2 are more closely related.

It does reconstruct protolanguage words, which I don't think is part of mass comparison per se.--JWB 23:49, 19 August 2005 (UTC)

"Amerind language and Finnish"

 * Thus, using this method, one critic was able to establish a correspondence between the proposed Amerind language and Finnish, and others were able to do so with Latin and many languages obviously not related to those of the Americas.

Was this cooking up individual false cognates, or a statistical analysis of a significant sample of words over a varied sample of languages? More information or a reference would be good.--JWB 03:35, 21 October 2005 (UTC)


 * Lacking a stochastic model, anecdotal examples like the French-Indonesian-Whatever are unscientific and prove nothing. Consider this statement: (A) there is a positive statistical correlation between the height of a man and his weight. Of course one can pick  a biased sample (or with biased measuring errors) that shows no correlation, or a negative correlation; but that is not a valid argument against  claim (A). Jorge Stolfi 16:56, 17 December 2005 (UTC)

Assessment
This section is very inaccurate. It does not deal at all with methodological that seem to be inherent to the method, does not quote any of those "many empirical tests", and claims that the only source of the problems is "unfamiliarity with statistical analysis".

The external link in the article explains some of the method's problems. To summarize a few problems:
 * It is pretty much arbitrary what constitutes a phonetic match (but this seems fixable)
 * It is completely arbitrary what constitutes a semantic match (and it does not seem to be fixable at all). This is not an issue of accuracy, like in just a random error, it's an issue of letting the person doing the analysis affect the results very significantly. It's very likely that 2 people doing the same analysis will get completely different results. It also makes "empirical validation" by checking whether traditional results are reproduces very dubious - for example the researcher may decide that words X and Y match because there's a regular shift between them, but X and Z don't, even though the distance between them is semantically and phonetically similar to one between X and Y.
 * If languages A, B are genetically related, a phonetic shift in either will reduce their phonetic similarity. A few such (completely regular) shifts, and the common origin is almost completely erased (assuming strict phonetic match). However, if language C recently borrowed a lot of words from A, they won't be affected by significant phonetic shifts in either. So it's very likely that C will be recorded as closer to A than B.


 * I believe that the last point is already covered in the article; see the caveat about borrowed words. As for the phonetic matching criterion, note that what matters is the relative similarity scores between language groups, not the absolute scores. Thus the phonetic matching criterion should be significant only if it was badly biased with respeect to systematic phonetic shifts in those languages. But for languages separated by millenia, the accumulation of many systematic shifts will tend to look like random drift. Similar arguments can be said about the arbitrariness of semantic matching. Jorge Stolfi 16:56, 17 December 2005 (UTC)

If someone wants to keep this section from removal, I'd suggest finding some references. It's very hard to defend it otherwise. Taw 16:59, 12 December 2005 (UTC)


 * It does not assert there is only one source of problems. It mentions one, and you can add others to the article text (as opposed to the current link where it is already available to readers). The normal Wikipedia procedure when there are conflicting major established points of view is to explain each and the arguments for and against, not delete.
 * Overall, I would say the current Assessment section has if anything an anti-MLC slant. The only other pro-MLC statements it makes are "basic soundness" (not defined, probably can be true or false depending on definition), reproduction of IE classification (one recent reference for mass lexical comparison on IE is ), and that Greenberg used the method in his work (it's no secret that Greenberg was an advocate of MLC, does anyone doubt he used it?). --JWB 20:52, 12 December 2005 (UTC)

The claims about Grenberg's African tree having been confirmed were obtained from net sources. Sorry, I don't have the refs; but isn't that a widely accepted point? Jorge Stolfi 21:19, 17 December 2005 (UTC)

Basically sound?
To call mass lexical comparison "basically sound" is a very bold claim.


 * Would you settle for "conceptually sound" or "theoretically sound"? (The criticisms seem to be all about the practical issue of when to define phonetic/semantic equivalence, not about the idea itself.) Jorge Stolfi 16:56, 17 December 2005 (UTC)

"Basically sound" with no elaboration doesn't say much and will continue to be a bone of contention. I'm in favor of rephrasing. I don't think "theoretically sound" itself is enough either though, since mass comparison has had concrete and widely accepted results (how about Greenberg's African classification?) as well as controversial ones.--JWB 03:52, 18 December 2005 (UTC)


 * It is definitely bold (in my view). While Greenberg's African proposals are now fairly widely accepted among specialists, his Amerindian claims are clearly not going to be accepted any time soon. There is hardly a single linguist working with Native American languages who buys into the theory. Part of the issue with the Amerindian proposal is that nearly a majority of the forms Greenberg cites as evidence are inaccurate in some way (incorrect or misleading gloss, mistranscribed, neglecting of known history, used to support two different comparison sets, etc.), and this conclusion has been demonstrated numerous times by different specialists in different investigations. This is one of the ultimate reasons that the Amerindian proposal is almost universally rejected by specialists (the other being, of course, that most linguists do not believe Mass Comparison itself to be scientific or accurate). A number of specialists with NA languages have called for Greenberg's Africa classification to be reassessed, in light of the spectacular errors in the Amerindian proposal.


 * "Theoretically" or "conceptually sound" worry me for this reason (if the method is successful in one attempt but fails in another, doubts have to be raised about the true soundness of the method). In any case, I disagree that it is theoretically or conceptually sound. Since Mass Comparison is unfalsifiable, it is in a sense not scientific, and the statistical foundation on which it is based is (I feel) fundamentally flawed. Mark Rosenfelder has two good articles on the subject:, . Of course, much of this is my subjective opionions, and I don't think the article should be completely biased against Mass Comparison. At the same time, though, I worry that portraying Mass Comparison as something that is agreed to be "theoretically sound" is a misrepresentation of the facts (at least in the sense that the majority of linguists do not agree with that assertion). I of course welcome any opposing opinions. --Whimemsz 21:14, 2 January 2006 (UTC)
 * Oh, and the claim in the article that the Amerind proposal is rejected "chiefly out of mistrust for the method" is false. As I said above, the huge numbers of errors and inaccuracies in the data are a large part of the reason that the proposal is so widely rejected by specialists. I think it should be removed or reworded (I'll get Campbell back soon, so I can cite my sources then). --Whimemsz 21:20, 2 January 2006 (UTC)

Well, as fas as I know, Greenberg & co. have carried out three major "experimental tests" of their method: (1) the Indo-European languages, (2) the languages of Africa, and (3) the languages of America. As an outsider, I am quite baffled by the virulence of the attacks against this method. From the score above, methinks it should deserve at least a measure of cautious respect &mdash; at least until the traditional method produces a reliable classification for (3) that is distinct from theirs. Instead, people keep inserting huge "STOP! READ NO FURTHER!" signs in this article, and prefixing "very extremely highly controversial" to every reference to the method in other articles --- as if reading a few sentences about this "sinful" method could thoroughly screw up the brains of innocent linguistics students.
 * In case (1) they got back the accepted solution; but granted that is no big deal because they knew it beforehand;
 * In case (2) they got a classification that was afterwards confirmed, at least in good part, by the traditional method;
 * In case (3) they got a classification, but it has neither been confirmed not rejected, because the traditional method still has to produce its own.

I also fail to see the logic in many of the criticisms. You say that Greenbrg's method is not falsifiable; to me it seems just as falsifiable as the traditional method. Perhaps what you mean is that its proponents remain unmoved, no matter how many errors have been pointed out in its application. But the whole point of a statistical method is to extract useful information from data that is contaminated by error &mdash; and all those semantic/phonetic mistakes of the list-makers are just another source of data error. In some disciplines that I am familiar with, one routinely and reliably extracts useful signals from data that is 99.9+% noise. (Of course there are conditions on how and when you can do that; chiefly you need only take a "large enough" amount of data.)  Biases in the errors is a legitimate worry, but only if the bias favor one tree over another. Uniform biases &mdash; that merely make all languages seem to be more related than they really are &mdash; are not a problem!

But it seems that, by and large, linguists are not used to think statistically. Rather they seem to classify their data as either "reliable/significant" or "unreliable/irrelevant", base all their deductions on the former, and ignore the latter. Now, I am not blaming them for that; given the difficulty of quantifying language phenomena, that is probably the best approach overall. But it does mean that to accept Greenberg's method they would have to radically change their established working paradigms. Could that be the source of all this animosity?

The virulence against Greenberg seems all the more baffling in view of the "establishment's" relatively bland rection to claims that, to me, seem much more preposterous, like the Sapir-Whorf hypothesis (see the Pirahã language and the recent report of the "Pirahã experiment" in Nature). Forgive me for this thought, but after seeing examples like those, I suspect that one can publish almost anything in linguistics &mdash; as long as one is careful not to contradict the currently fashionable theories, and tips the hat to the appropriate authorities.

Hm, come to think of it, that is not very different from what happens in suppposedly "exact" and "objective" fields, including my own. Sigh... All the best, Jorge Stolfi 22:46, 2 January 2006 (UTC)

In general, appreciation for Amerind is inversely proportional to one's training in linguistics and experience with Native American languages. I.e., most of the noisiest advocates for Amerind you'll see are people with no training in historical linguistics, or even linguistics in general, and definitely no expertise with the languages in question. They should take a hint from this, but I doubt they ever will, preferring to cast themselves as 'courageously fighting linguistic orthodoxy' or such. Anon., January 12, 2008 —Preceding unsigned comment added by 66.245.57.167 (talk) 01:34, 14 January 2008 (UTC)


 * Your point is irrelevant to Greenberg's work and other work of its type because your assumption that the method is statistical is false. The practictioners of mass lexical comparison do not perform any kind of statistical evaluation of the data. The method consists in its entirety of setting out lists of sets of words in different languages that resemble each other, to widely varying extents, in sound and meaning and saying: "Behold!". That's it. There is some work on truly statistical methods, which shows that they actually aren't very promising, but Greenberg, Ruhlen, and the like are not using these methods.Bill 18:05, 27 May 2006 (UTC)


 * If it makes you feel any better, I've never really believed a lot of what has been said about Piraha (or, to be more specific, I think that Daniel Everett has written in such a way as to make fairly "mundane" characteristics of the language seem more unusual. Example: there are, or were, lots of languages with no numbers above one or two). But then, I'm not a linguist.


 * Speaking of which, I'm obviously not a statistician either, so I'm not in a very good position to refute Greenberg there, although I nonetheless feel like the very idea of mass comparison (more or less randomly comparing data and deciding, based on guesswork/intuition, which words are related) is at best questionable on statistical ground.


 * By "unfalsifiable," I mean that negative evidence counts for nothing. If Ruhlen, for example, finds a word in some language like tuqi, meaning "finger," he would probably consider this a reflex of Proto-World *tik (I've moved on to Proto-World now, but the methods are basically the same). However, if he found another language where the word for "finger" is psei, he would not consider this evidence against Proto-World, but simply a language that had lost the reflex. (To some extent, this problem is present in using the comparative method as well, of course. My distrust of mass comparison doesn't mean I don't also distrust the comparative method...).


 * Whether he'll consider it evidence against Proto-World depends on the number and distribution of languages which have potential homologues of */tik/ vs those that lack them.
 * David Marjanović | david.marjanovic_at_gmx.at | 1:44 CET | 2006/2/17


 * The inaccuracies in the data are bad partly because they actually do end up skewing results in Greenberg's favor (I'm back to the Amerind proposal now). If an inaccuracy makes a form appear less closely related to other forms in the same set, Greenberg might decide it's not a cognate, and leave it out of the comparison. Whereas if an inaccuracy makes a form appear more closely related to other forms in the same set, Greenberg will surely leave it in, and the comparison will seem more impressive (I'm actually making quite a few assumptions/guesses here, but I think I'm right).


 * As you said, linguists aren't really alone in vicious in-fighting. Most scientific and semi-scientific disciplines tend to have several divisive controversies, with both sides (particularly the side with the majority, if there is a large difference in numbers on each side) bitterly denouncing the other side. Not that this makes it right...


 * Take care, --Whimemsz 23:46, 2 January 2006 (UTC)

Nakhleh et al on philogenetic trees for IE
The paper you quote is a very nice one. It uses statistical method, but a completely different one than the mass lexical comparison, and it's scientific standards are significantly higher than what we're suggesting in the article: However, they still get vastly different results from each set of seemingly valid parameters, most of them they found to be unacceptable. The worst (and significantly so) results were obtained from unscreened databases. Some notable quotations:
 * It analyzes phonological and morphological data, not only lexical data
 * It uses a precise definition of what constitutes a close match.
 * It carefully avoids problems, removing or weighting down data that is felt likely not to follow language's history (including most of the lexical data !)
 * It uses a wide range of well-established automatic methods for tree reconstruction.
 * The fact that G&A's method found such a tree from unscreened data is at least a strong argument that data should be screened when some methods are used.
 * Our observations thus strongly support the need for linguists to incorporate into cladistic analyses their own judgments about the relative reliability of different characters.
 * It seems possible that phylogenetic reconstruction methods are best suited to working out, in a maximally rigorous fashion, the consequences of linguists' judgments. Whether they can recover the actual history of a language family's diversification is a separate question.

So this paper, if anything, is an argument against mass lexical comparison, not for it. It may be considered an argument for statistical methods, but let's not equate the two. Taw 09:45, 17 December 2005 (UTC)


 * The paper is indeed nice and carefully done, but is their method really more "scientific" than Greenberg's? Their lexical character data too depends on semantic matching and arbitrary phonetic similarity criteria; it only seems precise after it has reduced to a numeric code.  Removal of "outliers" and of "problematic" variables is not a strong point, but a potential weakness (for which JG has been much criticised, it seems).  The use of historical information in the selection and weighting of the character database is potentially circular and weakens the conclusions.  Finally, the tree building methods is largely irrelevant, since they can be used with any similarity data; the issue is whether JG's similarity values are meaningful or not. This is not to criticise the paper, just to show that Greenberg's method is not that different, really. Jorge Stolfi 21:06, 17 December 2005 (UTC)


 * Mass comparison, as I understand it, is a more generic term. I don't see how it excludes Nakhleh et al. Taw, it almost seems that you are defining the term to only include sloppy work!
 * Before this article I've always seen "mass comparison", not "mass lexical comparison". The latter has less than 1/1000 as many Google hits, although many of the "mass comparison" hits are not about linguistics. I think the article should be about mass comparison without arbitrarily excluding the same methods on quantitative nonlexical data.--JWB 04:02, 18 December 2005 (UTC)


 * Mass comparison is phenetics. So are UPGMA and neighbor-joining. Only parsimony and Bayesian methods, among the methods used by Nakhleh et al., are cladistics (and so is probably "maximum compatibility" which I've never seen before; it seems to be almost the same as parsimony). Phenetics merely measures similarity; cladistics uses shared derived character states to reconstruct phylogeny. Phenetics can only give you a phylogenetic tree if there is little enough homoplasy in the data set -- not exactly a defensible assumption for a linguistic dataset! (Besides, it's one of the hypotheses we're trying to test by reconstructing the phylogeny.)
 * So, yes, at least two of their methods are (if correctly applied, anyway...) more scientific than that used by Greenberg. Greenberg's method, in turn, is (in principle...) a vast improvement over everything else simply because of the sheer amount of data it considers (...or can consider).
 * I just wonder what Nakhleh et al. used as the outgroup. Or how did they root their tree??? Did they simply assume that Anatolian must be the sister-group of all the rest?
 * David Marjanović | david.marjanovic_at_gmx.at | 1:41 CET | 2006/2/17

"find genetic relationships"?
The beginning of this article describes mass lexical comparison as a method to "find genetic relationships". As I understand it, MLC finds relationships which can either be genetic relation or the result of borrowing. Has anyone, including MLC proponents, ever asserted that similarity under MLC necessarily shows genetic relationship rather than a borrowing relationship?--JWB 00:13, 12 May 2006 (UTC)
 * You're right that, to the extent that MLC finds anything at all, it can't distinguish between similarities due to common descent and those due to language contact. Proponents of MLC, however, do indeed claim that it establishes common descent. Nothing in the method itself helps to do this. Basically, they claim: (a) that borrowing isn't very common with "basic vocabulary" and (b) that they remove the borrowings while examining the data.Bill 20:38, 11 May 2006 (UTC)

Acceptance
I have deleted the claim that the rejection of mass lexical comparison by most historical linguists is due to ignorance of statistics and to its revolutionary nature. The former is unsupported by any evidence. There is no evidence that the proponents of MLC have any statistical sophistication or that its critics have less. Indeed, the only serious mathematical discussions of MLC have come down squarely on the side of the critics. See the monograph by Ringe and the book by Kessler in the references. As to its revolutionary nature, MLC is not in any sense revolutionary. Rather, it is the old, pre-scientific approach that has gradually been supplanted by the comparative method. Historical linguists reject MLC not because it is new and unfamiliar but because it is old, familiar, and known not to work.Bill 20:54, 24 May 2006 (UTC)


 * Taking a look at Kessler's articles, he is certainly not against MLC, but is taking it even farther than Greenberg. He embraces and statistically validates some of the controversial features of the method criticized in the current version of this Wikipedia article, like using a small set of words (e.g. Swadesh list of 100 words works better than Swadesh 200), and considering only the point of articulation of the first phoneme of a word. And the algorithms he introduces are even more multilateral than Greenberg's, and add Monte Carlo methods. He criticizes some points of Greenberg's methodology, but in the context of refining the method, not throwing it out. He expresses doubts about the prospects for valid long-range comparison, but does not exclude the possibility that it can be achieved. He criticizes total rejection of long-range comparison or methods weaker than the comparative method:


 * Sheila Embleton notes that historical linguists have traditionally had a bias for stating only what is virtually certain. But there is also value in different kinds of claims, such as identifying the most probable of several alternatives, or in quantifying the likelihood of an uncertain possibility. It is in this spirit that most modern users of glottochronology work: They know it gives at best rough estimates, but when such limitations are openly acknowledged, rough estimates are arguably better than no estimate at all or estimates based on what Marisa Lohr calls “intuitive glottochronology”.


 * The current Acceptance section concentrates almost exclusively on debunking Greenberg's African classification, in enough detail that it would fit better in African languages or Joseph H. Greenberg. The article needs to add discussion of newer MLC work like Kessler's that addresses the criticisms, rather than claiming that MLC always produces wrong results. And citation of incorrect results rampant in early work like William Jones's needs to be qualified by the lack of basic glottochronology controls like limitation to a core vocabulary list or statistical analysis.--JWB 18:57, 28 May 2006 (UTC)
 * I agree that statistical methods like Kessler's deserve more consideration. In fact, as I write, I'm at the Banff International Research Station for Mathematical Innovation and Discovery, participating in a workshop on "Mathematical and computational approaches to linguistic phylogeny". Where I disagree is that I don't consider work like Greenberg's, and that includes almost everything that has been dubbed "mass lexical comparison", to fall into the same category as the statistical methods used by Kessler, or, for example, the earlier work by Robert Oswalt. In practice, "mass lexical comparison" consists of making lists of words that one considers, by unstated criteria, to resemble each other in sound and meaning, and saying "Behold!". There is no statistical evaluation of any kind. I think that it is crucial to bring out the totally unscientific character of this work, and I notice that quite a few people, including some who have written on this topic here on Wikipedia, have evidently never looked at the actual work and are under the misapprehension that Greenberg et al. use statistical techniques. What I would prefer is to have three articles: one on the comparative method, one on "mass lexical comparison", and one (that needs to be created) entitled something like "statistical methods for establishing linguistic affiliation". Note that such statistical methods do not haveto be restricted to the lexicon, which is another reason that I would prefer to keep them separate from mass lexical comparison. Bill 07:36, 29 May 2006 (UTC)

Fallacy

 * "find genetic relationships among languages in the remote past, beyond the limits of the traditional comparative method"

What limits? If you are talking about the famed 6000-year threshold, this is as weak an argument as Greenberg's. There is not a single major linguist nowadays that would spell out scientific reasons for the belief in a 6000 (or 3000, or 20000, for that matter) year limit for historical linguistic investigation through scientific comparative methods. Colin Renfrew even makes a case for this idea being a long-held "house of cards" (those are his words), which someone once suggested with no conclusive evidence and was just repeated all the way through, uncritically. Therfore, I don't think this should be (even through the current weasel words) in the text, let alone in the lead. For the moment, I'm removing it. 201.21.202.214 02:56, 19 September 2006 (UTC)
 * In principle, I think the idea that such a limit exists in linguistics is reasonable, because (unlike in biology) the average rate of change is pretty fast. But, firstly, if it were as low as 6000 or 10000 or 12000 years, why weren't IE and Afro-Asiatic extremely difficult to recognize? Why are they so plain obvious? Secondly, if such a limit exists and is as low as commonly claimed, what happens if we compare protolanguages that are themselves as old as the supposed limit, instead of comparing their descendants? This means the limit is pushed back another 6000 or 12000 years, right? Thirdly, "average rate of change" may not mean much; each word in a language changes at its own rate. In short, I agree with you, and I don't see any valid a priori reason to reject Nostratic or Dené-Caucasian.
 * David Marjanović | david.marjanovic_at_gmx.at | 13:32 CET | 2006/11/13

Inline citations
It would be helpful if this article made use of inline citations, so we know exactly who's making which criticism. --Ptcamn 09:38, 6 October 2006 (UTC)

Google Scholar search on title
"mass lexical comparison" 3 results

"mass comparison" (social sciences only) 202 results, but many still not linguistics

"mass comparison" author:greenberg 10 results (Some of these make clear that Greenberg himself used the term "mass comparison" and later "multilateral comparison")

I don't have any evidence that Greenberg himself used the term "mass lexical comparison". The few uses of this term are more recent and by critics. --JWB (talk) 04:35, 25 December 2007 (UTC)

assessment -> criticism
The "assessment" section was in fact a (highly inflammatory, POV) criticism section. I've changed the section name to reflect this and tried to make it less inflammatory. I removed a lot of verbiage that basically said nothing. In addition, I removed a specific assertion about a 1919 article by Radin that can hardly be relevant in the arguments over whether mass lexical comparison is valid due to its age. Benwing (talk) 04:11, 22 April 2008 (UTC)

Uralic
The Uralic pseudo-group is said to be "unanimously accepted" in the main article. Dr. Laszlo Maracz denies that Finnic is related to Ugric, and thus the whole of Uralic. See www.acronet.net —Preceding unsigned comment added by 86.143.5.71 (talk) 09:43, 26 April 2008 (UTC)

Same
Much the same applies to the Algonquian pseudo-group. The article on the same openly says "...out of convenience. Only Eastern Algonquian constitutes a true genetic subgroup." —Preceding unsigned comment added by 86.143.5.71 (talk) 09:51, 26 April 2008 (UTC)

MLC or MC
As noted in the Nakhleh discussion above, I believe Greenberg etc. used the term "mass comparison", not "mass lexical comparison", and used morphological and other comparisons in his work. I'm going to move this article to "Mass comparison" unless someone produces references showing "mass lexical comparison" is the dominant usage. --JWB (talk) 21:31, 20 July 2008 (UTC)