Language in the Americas

STANFORD UNIVERSITY PRESS

Contents

The Principles of Genetic Linguistic Classification

My purpose in this chapter is to discuss genetic classification; but I hope that the discussion will also help to explain and justify a deviation from what has become virtually a compulsory practice among American In-dianists: the use of sound correspondence tables and asterisked reconstructed forms.

In proceeding in this manner, this volume will resemble my first published classificational work, namely, that on African languages (Greenberg 1963). In that study I did not use a single asterisk or a single table representing a reconstructed sound system; and although I made occasional reference to particularly striking sound correspondences, these figured in no essential way as part of my method. There were, however, extensive lists of proposed etymologies, both lexical and grammatical, and shared grammatical irregularities. Yet it is reasonable to assert that this classification has won general acceptance and has become the basis for a considerable body of comparative work on African languages.

These aspects of my methodology earned me a fair share of criticisms, of course, even from those who accepted and built on my results in their own investigations. The following are representative quotations. The first are the views of William Welmers (1973: 5, 6, 15, respectively):

Greenberg has not, to be sure, demonstrated the existence of regular sound correspondences among all of the languages in any of the four language families he posits for Africa, although it has already been implied that such correspondences are the only real proof of genetic relationship. In fact, evidence that falls short of clear demonstration of regular phonetic correspondences may nevertheless be overwhelming. ... But the nature of the similar forms with similar meanings which Greenberg cites, and the number of them, is such that the fact of genetic relationship can be considered established. ... For all practical purposes the validity of the four families can be considered established.

Several years earlier, at a conference held at Aix, Luc Bouquiaux (1967: 156) made the following statement:

I do not assert that I accept the totality of his conclusions, but for the languages of the Jos Plateau in Nigeria which I know, my studies have in every instance confirmed the classification he has proposed. It is possible in fact that his metlod may not be absolutely correct in regard to the regularity of sound correspondenes, but I cannot but pay tribute to his intuition, which was later verified in every insance, although he often had at his disposal materials of very unequal value.

In a general review of the state of African linguistics, Paul Schachter (1971: 34) virtually stumbles on the correct solution in the fdlowing statement. (My italics indicate the decisive point.)

Certainly much more work of the kind begun by [J. M.] Stewart wil be needed before the same regularity of correspondences as that found within he Bantu family can be claimed for Niger-Congo as a whole, or for that matter fir any of its branches, none of which has to date been accorded the kind of schdarly scrutiny with which the Bantu languages have been favored. In the meantimc, however, it seems appropriate to ask what conclusions other than genetic relationship between Bantu and West African languages can be drawn from an objective examination of the data cited by Greenberg and his supporters; c.g. Greenberg's exensive lists of strikingly similar forms, with shared meanings, attested over the enire Niger-Congo areas, or the detailed morphophonemic similarities noted by Welmes.

But in an empirical science, how much more can be reaonably required than that the evidence be "overwhelming" and "the fact of genetic relationship [be] established" (Welmers), or that there be no 'other conclusion than genetic relationship" (Schachter) or that "the classfication ... proposed" be confirmed "in every instance" (Bouquiaux)?

Welmers' mention of demonstration, a term traditionally associated with Euclidean geometry, is appropriate in mathematics and logic, which were once described as consisting of "surprising tautologies.' The notion that regular sound correspondences can fittingly be called denonstrative in this sense, although this and similar terms have often beer used, will be shown in the course of this chapter to be illusory. As we shal see, what is in question is not just the nature of the truth exhibited by somd correspondences, but the still more basic question of what is meant by these and similar expressions, which are often used by linguists as though their meanings were self-evident.

There are indications that some investigators in the field of African languages have begun to realize that my work not only produced certain specific results, but also employed a revolution in methodology, as Edgar Gregersen (1977: 5), for example, has noted. Having made this point, Gregersen then quotes approvingly the following statement by Paul Newman (1970), a Chadic specialist: "The proof of genetic relationship does not depend on the demonstration of historical sound laws. Rather the discovery of sound laws and the reconstruction of linguistic history normally emerge from the careful comparison of languages already presumed to be related." Actually what is involved is not so much a revolution as a return, in certain essential respects, to an earlier point of view, as will be noted later in this chapter. In fact the Neo-Grammarians of the late nineteenth century, the very school that proclaimed the regularity of sound change as their central doctrine, never made the claims for it that have grown up since, and that have been accepted by many linguists as a virtually indisputable dogma — though never, I would add, either stated with clarity or reasonably proved, but simply taken for granted as axiomatic.

In discussing this doctrine critically, let me say at the outset, if it is not obvious, that my remarks are not intended as an attack on the validity of comparative linguistics or on the importance of undertaking reconstruction. Rather, the discussion is meant constructively as a way of taking first steps where the comparative method has not been applied for want of an assured basis in valid genetic classification.

I do not wish to claim that all of the points I shall raise in this discussion are original with me. Several have been made by others, and in the pages below their work will be duly noted. However, as far as I can see, the only persons who have thought along basically similar lines are Sydney Lamb, Aaron Dolgopolsky, and, to a lesser extent, the famous anthropologist Alfred Kroeber and his co-worker Roland Dixon.

Basically, the wrong question has been asked, namely, when are languages genetically related? Sometimes in fact it is phrased as follows: when are two languages genetically related? What should be asked is, how are languages to be classified genetically? Note that in all of the quotations above, the problem is stated in terms of relationship. As Lamb (1959: 33) notes, "To many linguists the classification of languages and the determination of relationships seem almost synonymous."

Consider this example. A linguist proposes the following classification for certain languages of Europe: (1) Swedish, Sicilian, and the Laconian dialect of Greek; (2) Norwegian and Provençal; (3) Bulgarian and Icelandic; and so forth. In every one of these groupings the languages are related, since they are all Indo-European. Moreover, we may credit our hypothetical classifier with caution, for if he proceeds in like fashion a large number of independent stocks will be proposed. What is absurd, of course, is that none of these groups is a valid genetic unit. By a valid genetic unit is meant a group at any level whose members are closer to each other genetically than to any form of speech outside the group. No doubt Bulgarian is related to Icelandic, but we are dealing here with a pseudo-entity from which strange cultural-historical conclusions would be drawn, and which does not constitute a reasonable unit for historical comparative investigation.

The concept of classification into valid genetic units in a hierarchy of various levels is a far richer notion than mere relationships. From such a detailed classification many statements of relationship of differing degrees can be deduced. Statements of relationship are thus mere consequences of classification, but not vice versa.

Note also that the above definition of valid genetic unit contained the phrase "closer to each other genetically than to any form of speech outside the group." The occurrence of "any" in this definition requires that one look exhaustively outside the group, since such external evidence is relevant to determining the validity of the group. Those, therefore, who focus on a limited group determined by accidents of expertise, and anywhere else they just happen to look, are anything but cautious. For what is more incautious than to disregard relevant evidence, as any trained historian will attest?

We may distinguish two kinds of lack of caution in these matters, asserting and denying. In the former, two languages or low-level groupings are compared to the exclusion of other languages at least equally closely related, as would happen if one compared Swedish and Albanian in isolation and asserted their relationship. Equally incautious is to deny a relationship while disregarding relevant evidence, as for example when an expert in a particular Hokan language who is skeptical of its Hokan affiliation, or indeed the existence of a Hokan group at all, looks at only one other Hokan language. A comparable case would be that of an Armenian specialist who, when told that Armenian is an Indo-European language, compared it only with English. With such a procedure, the specialist may well be overwhelmed by the differences, unable to evaluate the similarities for lack of a comprehensive comparative framework, and unaware of important pieces of evidence for Armenian being Indo-European because they happen not to appear in English, a point that will be developed in detail later in this chapter.

In light of the distinction between relationship and classification, the statement sometimes made that you cannot disprove the relationship of two languages becomes uninteresting. No doubt you cannot disprove that Nahuatl is related to Swahili, but you can disprove that Nahuatl is closer to Swahili than to Pima. It is to account for such comparative degrees of resemblance that one posits that Nahuatl and Pima must belong to some valid genetic group (in this instance Uto-Aztecan) that does not include Swahili.

We see that from this point of view the problems of subgrouping and classification are closer than has generally been realized. Indeed if all the languages of the world are related, the problems become identical. Classifying the languages of the world becomes simply a matter of subgrouping a single large stock.

But in subgrouping it is once more the distribution of similarities that counts. The significance of a particular similarity, in so far as it bears on classification, becomes apparent only when we know where else it is found. Put another way, the significance of distribution as the essential basis of historical inference is known to all historically oriented anthropologists, and language is merely a special case.

Note that all that has just been said is based on the notion of evaluating resemblances, and the point has sometimes been made that the notion of resemblance is vague. However, what is involved in classification is not the registering of a resemblance, but a noting of the comparative degree of resemblance. Is a form A more like B than it is like C? Given, for example, panlfanlezuk, who would hesitate? What is meant, moreover, by greater resemblance is diachronic resemblance, that is, the probability that A and B derived by changes from a common source, as compared with C's having derived from a common source at greater remove (e.g. fourlvierlcuatro) or from a different source altogether (e.g. hand/Hand/mana).

We may distinguish synchronic from diachronic resemblance even though they are enough alike that they can be largely equated in the heuristics of classification. Sounds and meanings by and large change to other sounds and meanings that are synchronically similar, e.g. the change from p to b, which involves a single feature difference. But s and h may be said to be diachronically similar because of the frequently attested change s > h, whereas in synchronic analysis they differ by a whole series of features. This example also illustrates another characteristic of diachronic resemblance, its frequent asymmetry, since h >s is not known to occur, whereas synchronically, by definition, s is as similar to h as h is to s.

Further, if we find three forms that all look very different from each other, no judgment of differential similarity is required. We base our classification on the strong predominance of similarities in one language, or set of languages, in comparison with another. There will no doubt be marginal cases, but even the most sophisticated techniques of the comparative method cannot decide all etymologies, as the reading of even a page or two of standard etymological dictionaries will show.

Also to be taken into account is the fact that as the number of languages ultimately known to belong to a grouping at some level increases, the precision of our judgment increases both in regard to the lower level of decisions, mere cognation, and in regard to at least some higher-level deductions regarding the shape and meaning of the source form.

Now we turn to the question of sound correspondences. Suppose there were a test that, when applied to two or more languages, always gave a definite answer. Let us suppose it is like a litmus test. The paper turns red when the languages are related, blue when they are not. Faced with, say, 1,000 languages in Africa, we begin to apply it. But even with pairwise comparison, there are 1,000 × 999/2, or 499,500, pairs we could choose. And even if such a test existed and gave valid results, the vast work of subgrouping would remain.

In fact, probably no one claims that we can devise a classification by regularity of sound correspondence, only that we can test hypotheses that have already been proposed. We therefore need a method of forming hypotheses. The number of ways of classifying n objects into one set, two sets, etc., up to n is called the partitions of n. Even without subgrouping, the number of partitions as a function of n increases astronomically with increasing values of n. The number of ways of classifying merely 20 languages is already 5,172 × 10, i.e. over 51 trillion. For 1,000 languages, of course, the number is far more staggering. How this is to be dealt with is discussed later in this chapter.

Those who have realized that as an initial step one must first choose some hypothesis in order to test it by regularity of correspondence maintain, then, that the comparative method is not a method of arriving at a classification, but a method of proving a classification already hypothesized. What is not taken into account is the truly astronomical number of possible classifications, as just noted. No method is given for choosing a hypothesis except "inspection" or perhaps intuition, as mentioned above by Bouquiaux.

Basically what I am denying is that there really are two separate steps. This possibility has been noted by some well-known philosophers of science. For example, in a discussion of Norwood R. Hanson's theories, Peter Achinstein (1977: 358) states: "Any of the reasons Hanson mentions for suggesting a hypothesis can also be, and often are, reasons for accepting it. Take Hanson's retroductive reasoning, the fact that a hypothesis offers a plausible explanation of the data can be a reason for accepting it once it has been suggested. There is no such thing as a logic of discovery as opposed to a logic of justification" (italics in the original).

Returning to our litmus-test analogy, we must, however, ask if any such test exists. Considering that such expressions as proof, demonstration, and certainty constantly recur in the literature, one can reasonably ask for a rigorous procedure. But in fact a variety of versions occur, usually not worked out in any detail, but alluded to as if generally understood, and equated in some fashion with the methodology of comparative linguistics as developed by the nineteenth-century Neo-Grammarians. I shall therefore set up a number of models that can be constructed on the basis of the sorts of statements that are commonly made in the literature. They seem to arrange themselves between two poles, an emphasis on regularity of correspondence at the one pole, and on the reconstruction of an ancestral or proto-language at the other.