Microsoft Word L 2-03-Greenberg-paper doc

mouth
a 
mouth
is no reason to call a 
nose
a 
nose
, though you 
will probably not call it a 
mouth
. This principle is of great importance in that for 
independent items the joint probability of accident becomes the product of their 
individual probabilities and hence is vanishingly small even with only a few instances. 
However, all items are not of equal weight. One consideration is length. Other things 
being equal, the longer the item the less likely it is to be accidental. Sound symbolism 
is another factor. The agreement of languages in having a word for the female parent 
such as 
mama
is obviously of relatively little weight. 
There is another sort of resemblance, on the other hand, which is of particularly 
great weight. Up to now we have simply talked about resemblances simultaneously 
involving sound and meaning. We may state this more exactly in the following form. 

---

The Methods and Purposes of Linguistic Genetic Classification
125 
The unit of interlingual comparison is the morpheme in the sense in which the term was 
used in American Structuralism. We are concerned with the morpheme as having in 
many instances a number of variant forms or allomorphs. Agreement in alternation 
among allomorphs is clearly of very great weight. The more irregular it is, the more 
powerful it becomes. The most powerful of all is agreement in suppletive alternation, 
where the allomorphs are derived from originally distinct morphemes. Thus the 
agreement of English with the other Germanic languages in the forms of the positive, 
comparative and superlative of 
good
is of such great weight that by itself it is sufficient 
to show that the Germanic languages are related to each other. However, it is not 
sufficient to show that the Germanic languages are a valid genetic group in the sense 
discussed earlier. The reason for asserting this is that the absence of this alternation is 
not sufficient in itself to prove that a language is not Germanic, since such irregularities 
are obviously the targets of analogical levelling. On the other hand, they are sometimes 
of such historical depth that they are evidence of groupings which exceed those of the 
level of Germanic in age. In the present case neither of these two strictures holds, but, 
of course, there is a vast amount of additional evidence to show that the Germanic 
languages are a valid genetic group at some level. Another way of saying this is that, at 
least taken in isolation from other resemblances, evidence of the type just discussed is 
useful for relationship rather than classification. 
In addition to the independence of each trait and their relative weighting, there is a 
third factor. This is the importance of the recurrence of similarities across more than 
two languages or language groups. Here as with trait independence there is a powerful 
probability factor. If the probability of an accidental resemblance between two 
languages is 
p
, then for three languages it is 
p
3/2
and, in general, for 
n
languages it is 
p
n
/2
. This rapidly becomes infinitesimal. Hence the agreement of a number of languages 
in a number of items, each logically independent but recurring over the same group of 
languages, provides the basic evidence for genetic grouping and is most easily brought 
into play by the technique of multilateral comparison. 
In distinguishing between relationship and classification we arrive at the second 
basic principle, one which is, I believe, the chief source of error at the present time. Our 
primary purpose is to classify languages genetically. This means that we seek to find 
valid genetic groups, that is, languages that are more closely related to each other than 
any is to any language outside the group. Thus Swedish, Albanian and Armenian are all 
related to each other, since they are Indo-European languages, but they do not 
constitute a valid genetic group at any level. Since classification is hierarchical, 
hypotheses of classification are much richer than those of relationship without level 
specified. From classifications we can deduce many hypotheses of relationship, but not 
vice versa. Thus, given a complete table of Indo-European classification, we can 

---

Joseph H. Greenberg
126 
deduce the statement above concerning the relationship of Swedish, Albanian, and 
Armenian, but from this fact alone we are not able to give a classification on any level. 
The situation does not change when we are dealing with deeper level 
classification. Thus a number of linguists have for a considerable period of time sought 
to show that Indo-European and Semitic are related. It is finally being realized (at least 
by some) that, since there is an obvious case for the greater resemblance of Semitic to 
Egyptian, Berber, Cushitic, and the Chadic groups, which form Afroasiatic, there is no 
point in comparing Indo-European with Semitic alone and the relationship, if it exists, 
must be with Afroasiatic as a whole. Most linguistic stocks do not have only two 
branches, and at an earlier period, in which isolation of human groups must have been 
greater than at the present period, this is even more likely to have been the case. Hence 
isolated hypotheses simply seeking to show that some language group is related to 
some other one, without bringing in a broader range of evidence to show that they form 
a valid grouping, is irrelevant. It is noteworthy that almost all hypotheses of this kind 
seek to connect some well-known or historically important family with another of the 
same sort or with a favorite language of the investigator, often his own. 
There is involved here a principle which we might call linguistic democracy: in 
forming hypotheses, all languages are of equal weight. In the late eighteenth and early 
nineteenth centuries there was a great reluctance among Hungarian linguists to admit 
that the languages closest to Hungarian were Vogul and Ostyak, although this obvious 
connection had been pointed out by a number of pioneer historical linguists. The most 
popular theory, at least among Hungarians, was that their language was related to 
Classical Greek. 
Another way of stating the foregoing considerations is that whenever we find a 
number of languages which resemble each other consistently, more than any resemble 
languages outside the group, we need an explanation of this obviously non-random 
phenomenon; and our explanation is that they are later developments from an earlier 
single ancestral or proto-language, as it is commonly called. When stated in this 
manner, it shows the intimate relation between the subgrouping and classification. In 
fact, if all the languages of the world are related, the problems become identical: the 
subgrouping of a single language family. A group stands out most easily (in regard to 
the types of resemblances just discussed) against the background of other groups which 
do not share the specific properties which mark out the group as such and distinguishes 
it from others. The best control against chance resemblances is not some fixed 
percentage but that furnished by other languages. 
The method just described is what has been called inspection and considered by 
many as “superficial”, in contrast to the comparative method which is based on regular 
sound correspondences. Actually, as we can see from the preceding sections, it is a very 

---

The Methods and Purposes of Linguistic Genetic Classification
127 
powerful method. Sometimes by inspection is meant merely pairwise comparison of 
languages. Clearly, this is not what is being advocated here. Moreover, in assessing 
resemblances, the existence of resemblant forms in a number of languages allows us to 
test much more adequately than with pairwise resemblances whether the forms have the 
hallmark of a valid etymology, namely that we can deduce, even if roughly, what the 
ancestral form must have been. 
In fact, there is no opposition between multilateral comparison and the 
comparative method. It is rather the first step in the comparative method itself. This is 
because, before we can start systematic comparison and reconstruction, we must know 
which languages to compare. The most that is claimed by the advocates of the 
comparative method in this restricted sense (that is, omitting the initial step of 
classification) is that it “proves” hypotheses of relationship, not that it produces the 
hypotheses that are to be proved. 
That the setting up of such hypotheses is a real problem can be shown from the 
following considerations. The possible ways of partitioning 
n
objects is a recursive 
function which grows at an enormous rate. For 25 languages it can be calculated that 
the number of classifications, without subgroupings, is of the order of 10
18
. For the 
hundreds or even thousands of languages with which we have to deal, the number of 
possible classifications is truly astronomical. Yet, if we simply examine a few basic 
words in all the languages of Europe, the correct classification into Indo-European, 
Finno-Ugric, and Basque fairly leaps to the eye by the time we have reached the second 
or third word, and along with this the universally accepted major subgroupings of Indo-
European. In actual practice what is used is essentially similar to the method of 
multilateral comparison, and it was utilized in making the basically correct 
classifications on which the comparative method was first employed. In fact, the 
essentials of this method were not worked out until at least a half century after the 
classifications were made, so they could not have been used in making them. Again, in 
Sub-Saharan Africa, Meinhof (1932), who did the first reconstruction of Proto-Bantu, 
had already decided (as had many before him) what a Bantu language was. He actually 
used only eight languages for his reconstruction. Later Guthrie (1967-71) also used a 
limited but larger sample of Bantu languages. There are literally hundreds of Bantu 
languages for which derivation from the reconstructed forms has never been carried 
out, and no one seriously doubts their Bantu affiliation. Yet one often encounters in the 
literature the statement that the genetic affiliation of a language is not proven until its 
derivation from a reconstructed proto-language has been demonstrated. 
Another case in point is Finno-Ugric and the larger Uralic family to which it 
belongs, along with Samoyed. The recognition of Finno-Ugric as a family preceded that 
of Indo-European (Sajnovics 1770), and even the most conservative today recognize 

---

Joseph H. Greenberg
128 
Finno-Ugric and the larger Uralic entity as valid. Yet in Szinnyei (1910), 140 years 
after the pioneer demonstration of Finno-Ugric, no completely reconstructed Proto-
Finno-Ugric forms are presented, although consonant correspondences are stated, a 
number of them problematic. As for the vowels, all that Szinnyei indicates is that a 
form contains either those with front or back harmony. Even this is uncertain in many 
instances. 
In Collinder (1960), for the first time to my knowledge, complete reconstructions 
are presented both for Finno-Ugric and the wider Uralic family. However they are 
preceded by the statement (p.405) that “it is a matter of course that in many instances 
the reconstruction is more uncertain than the etymology which it is based upon ... 
therefore the reader may put question marks