A Minimalist Program for Linguistics *

Bruce E. Nevin
Bolt Beranek & Newman/University of Pennsylvania

1. Introduction

Zellig Harris has an established place in what commonly passes for historiography of linguistics. In this now customary view, he was perhaps the most extreme among the so-called Post-Bloomfieldians who sought to devise discovery procedures whereby a grammar could be derived from distributional analysis of a corpus of utterances without reference to meaning. Taxonomic linguistics, as this has been called, would start with a corpus of phonetic records of speech and proceed by alternating steps of segmenting these records, classifying the segments according to their distributions relative to one another, and representing the data in terms of the resulting class-labels for another round of segmentation. It was scruples of empiricism (logical positivism) and behaviorist psychology, it is claimed, that forbade consideration of meaning.<1>

A responsible assessment of Zellig Harris' work cannot fairly begin until it has first been made obvious — not merely contended but made self-evident — that this now familiar picture is fundamentally wrong about Harris: not only that it asserts things that are not true, but also that it misses entirely what it was that Harris was after, and why he went after it in the way that he did.

Such an assessment, and even the prolegomenon to it that I have just proposed, is far too large a task to attempt here.<2> Instead, I will select one theme for discussion, namely, Harris' conception of meaning and linguistic information in respect to the methodology that he developed and demonstrated. I will touch on other themes only in passing, partly for the sake of context and partly to indicate the lines that such an investigation might take. (These tangential themes are summarized in Section 11.)

2. Motivation: Absence of an independent metalanguage

We will begin with the motivation for Harris' particular concerns regarding the methodology of linguistics. He observed that linguistics is set apart from the other sciences by a certain ineluctable characteristic of its subject matter. In any science (including linguistics) researchers communicate findings, propose interpretations, and reach agreements in an evolving consensus. To do this, they use language, especially but not at all exclusively the specialized science sublanguage<3> of their field. For any science, then, scientists use their shared language to reach agreements about a shared point of view on the subject matter of the science, prior to and apart from the science itself. But for linguistics the shared point of view (the agreements) attained by means of language cannot be prior to and apart from the subject matter of the science, which is after all language itself.<4> This, then, is the characterizing dilemma at the foundations of linguistics, and Harris' early recognition of it informs all his work as linguist and methodologist.

It is not clear to me how early Harris realized the import of this fact for linguistic theory and methodology. Munz (1972:269) provides some historical perspective on the observation that a language necessarily includes sentences stating anything that can be said about it metalinguistically ("that the language contains virtually unrestricted metalinguistic devices"), and suggests that the importance of this observation did not become clear to Harris until the 1960s. On internal evidence, some of which Munz cites, I am confident that the distinction between linguistic information and other types of meaning was reached quite early. This distinction was already nascent in writings of Bloomfield and in Sapir's trenchant explorations of individual participation in social pattern.<5>

Harris was by no means the first to observe that "one can form in any natural language various sentences that speak about that language (or about any other), about the combinations of parts in occurrences of the language, and the like" (1991:274). However, he took this observation further, recognizing that any grammar (definition, description, etc.) of a language consists of statements about the language that are made in the metalanguage contained in a natural language (or that depend upon statements in the metalanguage):

2.1 Form, information, and meaning

Harris' claim, then, is that linguistics as a science cannot have recourse to a metalanguage external to language and independent of it, and so can only define and describe the informational properties of language by employing those very properties to do so. This is not merely a methodological problem for linguistics, serious though that is; it is an important key to the essential nature of language. As we will see, it means that the information in an utterance is a function of redundancy in language and in the particular utterance. At first blush, this may seem trivially true, almost a tautology. But it has an important consequence as to what is a possible theory of linguistic semantics: the information in utterances cannot be a function of its being encoded into language from some prior representation of meaning, mental or otherwise.

In the conception taken for granted by many writers, language functions as or is used as a code, and that is how linguistic forms come to be correlated with meanings. The correlation itself, which is presupposed without question as something intuitively self evident, entails a distinction between forms and meanings as independent entities requiring correlation. This autonomy of linguistic form with respect to meaning Harris denies.<7>

It is not that distributional pattern on the one hand and meaning on the other are correlated somehow, as independent factors. What Harris calls linguistic information is indistinguishable from linguistic form because, indeed, it is constituted by it. This is necessarily so, for there is no means for representing the information in language that is not (or does not rest upon) language itself: no separate metalanguage.<8>

Linguistic information/form is conventional, socially shared. We as individuals apparently use this linguistic information to help organize our perceptions in associative memory and imagination. The preponderance of our perceptions are "meanings" that by contrast with linguistic information are subjective, idiosyncratic, notoriously not socially available or communicable in any reliable way. Meanings in this broad sense, which we perceive in all our activities and observations, are indeed correlated with linguistic information (identically, with linguistic form), but variably, idiosyncratically, without the reliability of social standardization that is the hallmark of linguistic information/form.

This is not to say that Harris proposes that there are two kinds of meanings, far less that he made such a distinction a priori and decided to investigate one of them, denominated linguistic information. Linguistic information is distinguished as a product of linguistic analysis. Other kinds of analysis, perhaps in conjunction with linguistic methods, might distinguish other kinds of meanings, as suggested for example for ethnographic analysis with respect to style and social or cultural meanings in Harris (1952c).

We presumably do correlate the socially standardized linguistic information in an utterance with the perceptions (current, imagined, and remembered) that constitute "meanings" for us in a more vivid and individualized sense. Certainly, we make just this presumption every time we use language with the intention of evoking such private perceptions in others, or with the intention of divining what perceptions of theirs may be relevant to our purposes. However, and despite our best intentions, we repeatedly find that the world of nonverbal perceptions is by and large much more subject to idiosyncratic variation and difference, between ourselves and others, and even within ourselves, than is the highly patterned domain of social conventions, especially those conventions constituting language. "Oh, is that what you meant", we say.<9>

Note carefully that this is not part of an argument for exclusion of meaning from linguistics on the grounds of it being too indeterminate and variable (as the "taxonomic linguistics" straw man would suggest). The point is more fundamental, and has to do with the relation between perception and thought on the one hand and language as a public institution on the other:

It is the constitutive capacity of language, creating a kind of information that cannot exist without language, that Harris is after. But to get at it, he cannot presuppose it, not only because of the characterizing dilemma of linguistics already noted, but also because linguistic information does not exist apart from linguistic form; it is simply not to be found in the world of real, imagined, and remembered nonlinguistic perceptions that constitute "meanings" in the broader sense for each of us idiosyncratically. The illusion that meaning or information can be considered separately and brought into correlation with the forms of a given language is an artifact of translation, either from the investigator's different native language or from some language-like system of mathematics or mathematical logic. In the first case, the investigator's native language constitutes in its structure the linguistic information that the investigator then takes as meanings supposedly given prior to the forms of the language being investigated. In the second case, the investigator can have learned (or created) the formal "language" and the meanings assigned to its elements and structures only by depending upon linguistic information in an antecedent natural language, and subsequent claims as to the meaningfulness of its formulaic expressions have the same dependency. In both cases we are talking about glosses and not meanings.

The informational properties of language — regularities of linguistic form — are notoriously not all obvious to a superficial inspection of speech. How can we enquire into them and disclose them without begging the question? It was to this question that Harris sought answers.

2.2 Consequences for a science of linguistics

The absence of an external metalanguage has important consequences, then, for linguistics. First, it shows the true motivation for Harris' distributional methodology:

In addition, it imposes a requirement on the results of linguistic analysis:

Only certain well-structured aspects of meaning, then, correlate directly with linguistic forms, and the methodological problem<13> is that one cannot presuppose the former (language-correlated structure in the world of perceived meanings) as a way of getting at the latter (the linguistic forms). The reason for this is that it is by way of the latter (language with its public, institutionalized constraints) that people constitute the former in socially standardized ways. Thus, the methodological problem and the aim of Harris' research are two sides of the same coin:

The condition that there is no external metalanguage independent of language itself, and the requirement that this condition imposes for a "least grammar", together determine the outline and direction of Harris's research, or at least the last four or five decades of it.

3. Socially constituted meaning

Two presuppositions of the preceding observations should be stated more explicitly. First, Harris shared with Bloomfield<14> and Sapir<15> the recognition that language is a social artifact. The `detached pattern' in language<16> is what makes it learnable as a social given, without which human cooperative action would come virtually to a standstill.

Secondly, I surmise that (like Bloomfield) Harris saw the formal analysis of grammar as an aspect of the study of semantics,<17> and with Sapir he saw that socially instituted form or pattern, especially that in language, constitutes aspects of meaning that could not exist without it. (See Harris 1951b, the long review article on Sapir 1949.) Harris saw, as Sapir had seen, that people have historically used language to build up, over generations, types of meaning that did not previously exist and in fact could not exist without language. This linguistic information is conventional, and it is objective to the degree that it is public, transmitted and shared by means of language, as distinct from the private, subjective world of perceptions apart from language. (Individuals in understanding utterances obviously do associate meanings of the latter sort with utterances—that is, with linguistic information—and in a way that is clearly related to it, but not in a socially standardized way.)

3.1 Contrast as social fact

The basis of language structure in social convention is seen first in the fundamental distinction between repetition and imitation. For Bloomfield, it was a "fundamental assumption of linguistics […] that in every speech-community some utterances are alike in form and meaning" (1933:78, emphasis in original). Members of the speech community perceive such utterances as repetitions. Yet they are not repetitions by virtue of imitation. How is repetition different from imitation?

Harris devised a simple operational test, the pair test for what is a repetition, so that the observation that some utterances in a language are "alike in form and meaning", i.e. are repetitions, is an assumption no longer. This test rests on the crucial, social fact that something conforming to a norm or convention may be repeated by producing another token of it, whereas something not participating in an institutionalized type-token relationship may only be imitated. (This differs too from the type-token relation seen in ordinary category perception, where we speak of another X, i.e. another member of category X, rather than a repetition of [category] X.)

Harris equates Bloomfield's `assumption' with the linguist's judgment, on the basis of "situation, meaning, and sounds", when the native speaker's perception that one utterance is a repetition of another cannot be determined or is not trusted (1951a:29fn1):

For Bloomfield, the linguist's determination of all repetition (and therefore of the phonemic contrasts of a language) was by assumption, based on the linguist's judgment of "situation, meaning, and sounds"; for Harris, only a residuum of cases was, requiring only passing notice in a footnote. Harris' approach relies upon the social basis for repetition and contrast as a learned norm institutionalized among native speakers, rather than making it a necessary assumption by the linguist. See also Bloomfield (1933: 77, 128).

Harris describes the pair test as follows:<18>

A comment on the last sentence of this passage may be required: although neither the linguist nor the native speakers use the meanings of a pair of forms to judge whether they contrast or are repetitions, the perception of repetition vs. contrast is itself the irreducible, elemental, least datum of linguistic semantics: the perception of those least differences that make a difference in the language.<19>

In any case, it is important to recognize that meaning was seen to be problematic because it was difficult to manage, and not because of Empiricist strictures against mentalist concepts. As Harris put it (1951a:20, similarly at 7fn4):

3.2 Deriving new elements from redundancy

The pair test is not itself a distributional test, but rather an operational definition of the least elements, the phonemic contrasts. "The fundamental data of descriptive linguistics are … the distinctions and equivalences among utterances and parts of utterances" (1951a:33). It is because the contrasts are the fundamental (socially determined) elements of language that subsequent distributional tests "reduce to" the pair test, as Harris says.

Distributional methods in linguistics identify and analyze patterns of redundancy among the elements of language. The distributional methods themselves rest upon the correlation of redundancy with information shown by Shannon and Weaver in the mathematical theory of communication (Shannon 1949).<20> Distributional restrictions among elements are used to define new elements that are less restricted.

By defining more general (`higher-level') elements in terms of restrictions on prior (`lower-level') elements, those restrictions are removed from the representation of utterances. The restrictions are not removed entirely, but only moved to the definitions of the new, more freely combining elements. The restrictions that define the new elements (phonemes out of contrasts, morphemes out of phonemes, constructions out of morphemes) reflect the contribution of those new elements to the information in utterances. The total redundancy in an utterance is made up of a hierarchy of contributory redundancies. Redundancy among entities on one level constitutes entities of a different type on the next level of generality,<21> which in turn bear their own redundancies. This, in part, is how language creates information.

Beginning in the 1940s, Harris asked: Given a representation of utterances in terms of linguistic elements at a given level of generality, what are the redundancies found among those elements? Obviously, not all possible combinations of phonemes occur.<22> Only certain sequences constitute morphemes and words. But there is greater freedom of combinability between words, and between morphemes, than there is within them.<23> This difference in combinatorial possibilities gives a distributional basis (defined by a stochastic process) for the recognition of morpheme boundaries (Harris 1954:13, 1955, 1967, 1968:24-28).<24>

Given the morpheme boundaries,<25> we observe that not all possible sequences of morphemes occur (or have the same likelihood of occurring). This is for by now familiar information-theoretic reasons: if all combinations of entities in an ensemble are equiprobable, there can be no information in one combination as distinct from another. Determining which combinations can occur is "the central problem of structural linguistics" (Harris 1968:13):<26>

4. Tools of analysis

In the course of work on this "central problem", Harris invented an astonishing range of methods for analysis of language, including an algebraic treatment of constituent structure, with provisions for its inherent weaknesses (1946, 1963),<27> string analysis as a complementary tool compensating for those and other weaknesses of immediate constituent analysis<28> (1959, 1961, 1962), transformational analysis (1952a, 1952b, 1956, 1957, 1965), discourse analysis (1952a, 1952b, 1952c, 1989), and sublanguage analysis (1982b, 1985, 1989).<29> It is important to recognize that Harris saw these as tools of linguistic analysis, not as competing theories of language nor even as bearing directly on the question of the logical form of a theory of language. In a well known comparison of string analysis, immediate-constituent analysis (phrase-structure grammar), and transformational analysis, he wrote (1965:238—239):

The impulse to hypostatize as Language what are really only artifacts or reflections of the tools used for investigating utterances has always been with us. Harris resisted the impulse to rush to closure because he had larger ends in view.<34> Each of these approaches revealed only some of the characteristic (informational) properties of language; each was incapable of coming to closure over all of these properties, or could be forced to closure only by adding to the redundancy of the description by various ad hoc adjustments and exceptions for difficult cases, which would obscure the essential redundancy that constitutes the information in discourses.

5. Distributionalism

We will turn now to a closer examination of the ways in which Harris extended and refined the methods of linguistic analysis pioneered by Bloomfield and Sapir, which had come to be called distributional analysis.<35> As we have seen, this is a "methodological approach, of defining more freely combining new elements on the basis of occurrence-restrictions of old elements" (Harris 1970:v). Distributional methods disclose, as Harris said in respect to Sapir's way of working, a `detached' pattern in language. In part, this means (as Bloomfield and Sapir also had seen) that factors external to language, such as might be proposed in psychology, cannot legitimately be invoked as explanatory principles for the patterning in language, though such factors may have interpretative relevance to the results of formal analysis.<36>

Harris knew that such lines of `explanation', if taken up prematurely, could only interfere with the work proper to linguistics:

This view contrasts sharply with the expectation of many linguists that each discovered bit of regularity is an outward sign of some inward and innate explanatory principle, and the perception that research is worthy of pursuit or support only insofar as it advances or refutes one proposed explanatory principle or another. This way of working fosters fragmentary and mutually inconsistent results. By this I refer not to disagreements among linguists with alternative analyses, but rather to fragments of grammatical description that cannot be reconciled with each other for complete coverage of any single language. See also, for example, Harris (1954:13):

A footnote in this passage touches (in 1954) on the dilemma due to the lack of an independent metalanguage:

5.1 Practical distributionalism

Munz (1972) distinguishes three types or degrees of distributionalism: practical, theoretical, and interpretive. As we noted at the outset of this paper, it is commonly alleged that Harris was trying to work out formal analytic techniques that could be applied mechanically (using a computer) to a corpus of linguistic data to produce a grammar.<37> This practical distributionalism is clearly a straw man, at least as regards Harris' work.

The existence of computer programs that demonstrate or test aspects of Harris' work, even the fact that certain aspects of language structure can be (partially) determined by such applications, thus does not gainsay the essential point that practical distributionalism was not his aim.

The situation is somewhat different with sublanguage analysis, especially for sublanguages of science, e.g., Hirschman (1986). Because the metalanguage for sublanguage grammar is external to the sublanguage (being the same metalanguage as for the language as a whole), practical discovery procedures are possible that cannot work over unrestricted language domains. See Harris (1991 Chap 10) for discussion, and Grishman et al. (1986) for an example of work in progress.

Probably the closest thing to a practical discovery procedure in Harris' oeuvre is his procedure for identifying morpheme boundaries (1955). What resulted from a computational test using very large word lists (from several large on-line dictionaries in standard English orthography) was a good determination of most morpheme boundaries, with a residuum to be resolved by other distributional methods (Harris 1967).

5.2 Theoretical distributionalism

But Harris did not regard the phoneme-to-morpheme prodecure as a practical discovery procedure:

In Harris' own view, then, this procedure amounts at most to what Munz (1972) terms "theoretical distributionalism". This is the notion that a grammar would result if one used proposed procedures on a sufficiently large corpus for a sufficiently long time, though no one would in fact do so in practice. And this is the ground on which the strongest case might be built for a claim that Harris' aim was the devising of practical discovery procedures.

5.3 Interpretive distributionalism

This leaves for virtually all of Harris' work only `interpretive distributionalism': the view that rules of grammar, however determined, are statements about distribution and are falsifiable with respect to distribution. Munz suggests that this view is merely a minimal requirement for any linguist to be taken seriously. However, because there is no external metalanguage, Munz's interpretive distributionalism must be understood in stronger terms: distinctions among linguistic elements (distinguished for a language in a grammar) must be grounded in the distributional redundancy of previously established elements, resting ultimately on the fundamental social definition of phonological contrasts.

In practice, the linguist working in the field leans on a previously learned language, just as any other scientist does for whatever science domain. The pitfalls of doing so for language description are well known, at least in principle, as is the value of distributional methods as safeguards against projecting inappropriate linguistic categories from one's understanding of one language structure onto that of another (see e.g. 1951a:2-3). Also reasonably familiar is the fact that distributional methods open up for consideration alternatives that might otherwise be neglected (e.g. 1951a:8fn7) and give a principled basis for fundamental decisions (e.g. 1951a:32fn8). One form of this is illustrated by Harris' discussion (1954:9-11) of a Cherokee paradigm, where the fact that some elements have no clear English glosses is irrelevant, given distributional support. The converse situation is illustrated by the fact (1951a:193-194, 1954, and elsewhere) that the sl-, gl-, fl- words in English are not analyzable as e.g. an sl- prefix plus stem, absent distributional support, despite the correlation of identifiable meanings with word sets like slide, slither, and slip. The sometimes startling originality of Harris' formulations of sources and derivations in operator grammar was enabled, one suspects, by his returning again and again to the distributional facts without prejudice as to their expected interpretation.<39>

5.4 Evaluation criteria

Interpretive distributionalism thus posits a basis or criterion by which to evaluate and justify alternative proposed analyses. Absent an external metalanguage, this criterion is as we have seen a necessary one, but given the non-uniqueness of distributionally justified analyses it is not a sufficient one for determining the `best' analysis. Harris repeatedly states clearly that additional criteria are needed of two sorts, which we may here call interior and exterior criteria.

Interior criteria concern relations between one part of the grammar or linguistic description and another. One of the distributionally valid analyses on a given level within the grammar (e.g. phonological) may be more advangageous on the next level (e.g. in determining and specifying morphemes and morphophonemics).

Exterior criteria concern the purposes that one wishes the description to serve. For example, a representation of the phonological contrasts that is optimum for studies of phonological processes may not be the most convenient representation for historical comparison and reconstruction. What is constant under the non-uniqueness of phonological descriptions is the contrasts, which are based in socially established perceptions of what is a repetition.

What is determined here is not `the' grammar of `the' language in some absolute sense,<40> but the optimum description for particular ends, such as fitting into a particular framework in philosophy or psychology to which, for reasons outside of linguistics, one may be committed.

6. Imperfect coverage

Distributional methods avowedly do not provide complete or ideal coverage (1954:12). One reason is the non-uniqueness already discussed. (We should note here that non-uniqueness is of course not at all limited to phonemics. Harris (1951a) took considerable pains to indicate alternative ways of setting up elements at all levels of description.)

An even more fundamental reason is that language is not well defined, a characteristic that has been much discussed in terms of degrees of grammaticality, degrees of acceptability, `squishes', and so on. It may seem that such subtleties are beyond the capacity of distributionalism. It is important, then, to see how Harris addressed these problems without having his theory depend upon a putative independent metalanguage concealed behind elaborated judgments of grammaticality or meaning.

6.1 Graded membership

To accomodate this essential characteristic of language, that the set of sentences is not well defined, Harris refined the definitions of scope and domain used in distributional analysis of transformations. In the earlier definition of transformation (1952a, 1956, 1957), if any word-selection that satisfies one sentence form<41> also satisfies a second sentence form, then a transformation exists as a mapping between the two sets of satisfier-sentences (or, at a later stage of analysis, from one set into the other, framing a derivational sequence).

But in the criterion for transformation used first in Harris (1965)—though anticipated as a possible requirement in (1957)—satisfiers of two sentence forms are graded as to acceptability. If the relative acceptability of a pair of sentences satisfying form A is not reversed for the corresponding pair satisfying form B (for any pairs of satisfiers), the two sentence forms are transforms.<42> This criterion remains especially important for deciding marginal cases as the gross transformations are analyzed into elementary increments (ultimately operators) and operations (linearization, transpositions, and morphophonemic reductions).

There are then two points at which Harris' methods depend in principle upon native intuitions. This extension of distributional methodology does indeed call upon intuitive judgments of native speakers, but in a constrained, pairwise manner, and without reference to a prior metalanguage specifying meanings. It is not difficult to draw a parallel between this criterion for transformation and the pair test for repetition, which determines phonemic contrast. (To forestall misconstrual, let it be said once again that even a non-native linguist relies in practice upon intuitive grasp of the language at every point of practical linguistic analysis.)

6.2 From exceptions to extensions

Another reason that distributional methods do not give ideal coverage is the great number of irregularities in any language, exceptions to the regular patterns that they treat so well. Examples include asyntactic utterances, such as "Fire!",<43> and "the mass of idiomatic and quasi-idiomatic expressions in language" (1965:239).

In the work of wrestling the mappings in the set of sentences first into elementary transformations and then into operator-argument dependencies plus reductions, the residuum of exceptions come to be handled in a natural way as extensions of established relations and reductions. The advantages for this of distributional methods was clear even for morphological analysis (e.g. 1951a:8fn7):

It is in the analysis of syntax and information structures, however, that the power of distributional methods to account for exceptions by analogy to regularities is most clearly seen:

This is the principle of regularization or normalization that is familiar in phonology and in morphophonemics. Harris (1951b) describes this as a characteristic of Sapir's way of working. Here he applies the same method, termed `extended morphophonemics', to syntax. (See, for example, Harris [1965:239fn7], on exceptions, quoted earlier.) The basis for the regularization by analogic extension lies in the mandate for least grammar:

The analysis of complex transformations into elementary sentence-differences and thence into the sentence-forming steps of operator grammar (word entry, linearization of word dependencies into word strings, and reductions of word shape), is the natural and necessary consequence of acting on that mandate:

Some of the derivational sources and some of the intermediate steps in these derivations are sentence forms whose satisfiers are all of low or dubious acceptability. More often, sources for individual words or affixes are not attested or naturally accepted in the language today. These, Harris marks with a dagger, and he allows them (exploiting the ill-definition of the set of sentences) only if they are reachable in the derivation by steps that are well attested between fully acceptable sentences in other, different derivations:

Harris' image used in a seminar (in discussion of Harris 1969) showed two concentric circles: the regular structure due to operator-argument dependencies, with its close correspondence of form with information, and the extension of this by reductions into the larger set of more idiomatic paraphrases. Superposed over this was a wavy line, approximating the larger circle, showing the fuzzy boundary of questionable sentences and "infrasentences" at the margins of language. Metaphor and other analogic processes result in uncertainty and continual adjustment of these boundaries.

Harris' treatment of exceptions by analogic extension of regularities in the language provides (and rests upon) insights into processes of language change and variation.

It is important to realize that what is involved here is not merely a capacity of the theory (or of a grammar based in the theory) to accomodate marginal data, but rather a capacity of language which Harris' theory describes. The reality of it in language can be seen in processes of language change and in hearers' ready accomodation to language variation, and the relevance for synchronic linguistic description we have just seen:

Furthermore, this capacity of language (that is, of its speakers) to accomodate even the irregularities of language as extensions of its regularities, in a manner very like folk etymology, is necessary for language to be learnable and socially transmissible:

7. `Taxonomic linguistics'

So far in this discussion of Harris' distributionalism, the relation of distributional patterning to meaning has naturally been in the foreground, as indeed it was for Harris. As subsidiary topics, we have touched upon the canard about practical discovery procedures and about Harris' methods being restricted to a fixed corpus. In what has become the customary representation of Harris' work, it is these misapprehensions that have the greater prominence.

To these three forms of distributionalism, then, we may add a familiar straw man, Taxonomic Linguistics: "Its most basic techniques for arriving at […] a grammar are segmentation and classification" (Chomsky 1971:65). "It was believed quite literally that starting with a corpus of utterances, a grammar could be constructed through successive segmentations and classifications" (Newmeyer 1980:6-7).

Setting aside the allegations about discovery procedures, which we have already considered, this description applies superficially to immediate constituent analysis, in terms of which still-prevalent conceptions of syntactic analysis were formed.<44> However, Harris' distributional methods proceed not by segmentation of longer stretches and classification of the resulting segments, but by "defining more freely combining new elements on the basis of occurrence-restrictions of old elements" (1970:v).

Thus, the pair test for repetition replaces the continua of speech sounds by the relation of contrast.<45> Judgments as to the relative acceptability of homomorphic utterances—the use of acceptability-gradings as the criterion for transformation given beginning in Harris (1965) was noted earlier—replace an attempt to distinguish categorically what is in the language and what is not.<46> Relations of stochastic dependency among phonemes (that is, among the phonemic contrasts, however represented) define morpheme boundaries, replacing a distinguishing of them in respect to their glosses in some prior language that has been given the role of metalanguage. Relations of dependency between operators and their argument words, under reduction, replace the pseudo-hierarchies of phrase classes of constituent grammar and the string types of adjunction grammar.

8. Summary of Harris' methodology

The chapter on method in Harris' last book (1991:30-49) may be his best summary of his distributional approach. Here, he states unequivocally that the motivation of his methods is the creation of a coherent, consistent, and well founded theory shaped by the character of language itself, rather than being shaped by prior psychological or philosophical commitments. I can quote only part of this summary here:

Harris' footnote (in the last paragraph, just after the reference to section 2.3), succinctly addresses the questions of discovery procedures, evaluation procedures, and some aspects of language universals:<48>

9. Consistency of research into linguistic information

The aim of research into distributional regularities in language, then, was not to expunge meaning from grammar, but to disclose it in grammar. And Harris' interest in identifying the structures in language that carry or constitute information is not a recent change of heart evinced only in his last books, as some might suppose.

For obvious example, this interest is explicitly manifest in the early work on discourse analysis. It is perhaps less obvious in string analysis,<49> but its relevance becomes apparent in consideration of how string analysis is carried out. In the practice of analyzing linguisic data, one repeatedly asks what least subpart one may remove from a given utterance without destroying its acceptability.<50> What remains is an elementary sentence called the center or center string; the parts excised are adjuncts, either to the center or to other, previously entering adjuncts. Each substring makes its own identified contribution to the information in the sentence.

String grammar was first developed as a framework for computable syntactic analysis. A string framework for transformational analysis<51> has lent itself well to applications of discourse analysis in which structured databases are generated from free-form sublanguage texts. (This work is partially summarized in Sager 1975; see also Sager (1986).) It is especially relevant to the string approach that members of a `word' class in a sublanguage grammar in many cases comprises several morphemes and words in a general-coverage grammar or in the grammar of a different sublanguage. For example, the phrase the beating of the heart is a member of the SYMPTOM class in a sublanguage of pharmacology (in, e.g., Digitalis affects the beating of the heart), but in a sublanguage of the logically and epistemically prior science physiology this phrase is analyzed further, with heart being a member of the BODY-PART class and beating a member of the ACTION class (see Harris 1982b for discussion).

10. Sublanguage analysis and science languages

We will consider last Harris' pioneering of sublanguage analysis in the light of two of his longstanding interests (shared with Bloomfield and Sapir): in auxiliary languages for international cooperation, and in the application of linguistics to the advancement of science. Informally, a sublanguage may be thought of as a form of language specialized for a particular subject matter, especially a technical subject matter. Sublanguage grammar is freed from the proscription against an external metalanguage. This is because sublanguage terms may be defined in ordinary language terms, and the metalanguage for the language as a whole<52> is indeed external to and prior to the sublanguage grammar that it describes. This encapsulation gives rise to capacities and characteristics of sublanguage grammar that are not otherwise available for grammar—for example, the capacity to distinguish nonsense in a well-defined way, and even certain types of practical discovery procedures for sublanguage analysis (e.g. Grishman et al. (1986)).

As it turns out, the information structures in the sublanguage of a field of science closely parallel and (apparently) help to constitute the subject matter of that field: changes in scientists' understanding of the subject matter of a science subfield appear to be closely parallelled by concurrent changes in the structure of the sublanguage for the subfield. Indeed, the concurrency suggests that the latter may in part support, facilitate, or enable the former (Harris et al. 1989). The task of translation is far simpler and more direct in science sublanguages than in the general case, and when sublanguage grammar of a given field is compared cross-linguistically (e.g. between English and French or German), the same structures are found, and the same changes of structure concurrent with changes in the field. This indicates that something new has developed here: not a sublanguage of any one language, but a science language.

A science language shows certain sentence types whose structures would emerge outside the science language grammar only in discourse analysis. In contrast with the openness of ordinary vocabulary, its vocabulary is effectively closed at any given time. Selection is binary (yes-no), as distinct from the graded selection of ordinary language, and it is this that enables the clear demarcation of nonsense in science. New and more complex argument-dependencies may be emerging in science languages, and other differences and new capacities with respect to ordinary language may be disclosed by further research.

11. Conclusion

In this brief survey of selected aspects of Harris' work I have tried to show that the commonplace view of Harris in the `Neo-Bloomfieldian' stereotype is badly mistaken. According to this standard conception, he sought to devise discovery procedures whereby a grammar could be derived from distributional analysis of a corpus of utterances without reference to meaning. His avoidance of meaning, it is ordinarily claimed, was motivated by empiricist (logical positivist) and behaviorist principles that forbade reference to "mentalist" concepts. Against this view, I have argued that Harris at first avoided reliance on judgments of meaning only because they are too difficult to control (e.g., "struggling with exact meanings (1951a:31fn4)), and that he later showed why meanings are necessarily difficult to control in any way that is relevant for linguistics, since (except for sublanguage grammar) there is no means outside of language itself by which to specify them (no external metalanguage). More generally, I have shown how Harris has from an early stage been engaged in the study of linguistic semantics, and that he achieved results that are of very great importance for linguistics and for the advancement of science in general.

I have barely indicated some of the other threads that must be taken up in any responsible assessment of Harris' work, such as his emphasis on social convention as distinct from physical or biological determinants of individual psychology, his perspective on language variation and change, and his suggestions regarding the origins of language, both in language acquisition by individuals and in the evolution of the species. Other themes have suggested themselves just in the course of this writing, including ongoing work on sublanguage analysis, particularly the study of argumentation and proof in languages of science and of the emergence of new capacities not found in ordinary language; Harris' conception of the proper conduct of scientific inquiry, the role of methodology, the range of methodological questions he investigated and his way of working with them; implications that may be drawn from the essential properties of language that Harris identified as making possible a mathematical treatment; the application of Harris' principle of least description to other theories of language as a "test of relevance" (though Harris would never have undertaken or encouraged such comparisons); the use of Harris' theory of linguistic information in connection with ethnographic, sociological, psychological, stylistic, literary, and other studies of language use; the origins of the autonomy hypothesis, especially in relation to the descriptive incapacities and the enforced abstractness of phrase-structure based formalisms; the potential for an international language for scientific cooperation. Countless other lines of research are possible, as with any paradigm of work in science. One might consider in terms of operator grammar issues that have been problematic in Generativist theory at one time or another, as for example unbounded dependencies (Nevin, MS) or quantifier scope. Harris himself expressed interest in sign language and poetics (p.c. following one of the Bampton lectures at Columbia Univ. in 1986).

I have touched on only a few of the misconstruals of Harris' earlier work by his contemporaries (which have often been taken as definitive) and in more recent writings about American structuralism. Among the direct consequences of this reassessment of Harris' views is the conclusion that he cannot be held responsible for the doctrine of the autonomy of linguistic form. With Bloomfield and Sapir, he saw that linguistic form (pattern, configuration) is identical with that aspect of meaning which is most transmissible precisely because it is socially constituted with language.

Munz (1972), followed by Hymes & Fought (1981), suggests that Harris' views gradually shifted from distributionalism to a concern with meaning during the 1950s and 1960s. One result of the examination undertaken here should be the recognition that the motivation of Harris' distributionalism from an early stage was semantic, namely, the investigation of the meanings that are constituted by linguistic form, and that in this he carried forward like concerns of Bloomfield and Sapir. "It was all there from the beginning; there was no `shifting from distributionalism to meaning', only the constant and unexpected restating and reshaping. After all, the early transformations (with Chomsky and Lukoff) were precisely aimed at the entities whose distribution is appropriate to syntax (and semantics). That is why the work was so new and dazzling" (Hoenigswald, p.c.). This continuity has been obscured by the emergence of the `autonomy hypothesis' in generativist theory (really, several distinct hypotheses—see Ryckman 1986:179-196), and the rhetoric of standing on the shoulders of disavowed predecessors. That too, however, is a thread to be unravelled on another occasion.

Author's address:

Bruce E. Nevin
RR 2 Box 351A
Edgartown, MA 02539-9802 U.S.A.
bnevin@cisco.com

References

Notes

* I am grateful for the helpful comments of Professors Michael Gottfried, Henry Hoenigswald, & E. F. Konrad Koerner on earlier drafts of this paper. Responsibility for errors of fact and interpretation rests, of course, with me.

** As of 1996, Cisco Systems.

1. This characterization is so familiar and well established as scarcely to need citation. It has pervaded even the literature of popularization, for example, Campbell (1972:171):

Campbell gives no source for this view, which evidently appears to him to be part of the unquestioned background of his reading and conversations with linguists (he does cite conversations with Ray Jackendoff).

The pervasiveness of this caricature in the literature may be seen in R. A. Harris (1993), e.g., "a cornerstone of transformational research right back to Harris" (p. 428), as though he were only an historical figure of the 1950s whose work ended there; accounting Harris among the "Post-Bloomfieldians, from the linguistics boneyard" (p. 429); "the asemantic transformational granddaddy, Zellig Harris" (ibid.).

2. This task must be taken up by others better situated than I in terms both of knowledge and resources. For an excellent beginning, see Ryckman (1986), on which I have depended heavily for its masterful survey of the literature. Anyone could verify claims made here about Harris by studying what he has written on its own terms, without undue preconception. Ideally, the present brief exploration might provide entrée into what "its own terms" are, on which Harris' oeuvre must be understood, and might indicate some of the hedge of preconceptions surrounding his work that must be cleared away before it can be rightly understood.

3. For the present, a science sublanguage may be thought of informally as a specialized technical jargon. For more precise definition and discussion of sublanguage grammar and of science sublanguages in particular, see Harris et al. (1989), Harris (1982b, 1988, 1991), Grishman & Kittredge (1986), Kittredge & Lehrberger (1982), and Sager (1986).

4. This is of course, as said above, an evolving consensus, and as we shall see the specialized forms of language used in a science appear to have a constitutive role in the development of the subject matter of the science.

For any science but linguistics, one may use ordinary language as a metalanguage to define the objects and constructions of the science. For example, the talk and writing in which endocrinologists formulate, present, and discuss findings in their field is clearly within the language that they share with non-endocrinologists, however specialized for their work. The talk and writing in which linguists formulate, present, and discuss grammar is likewise necessarily part of the shared "background vernacular" of language. It goes without saying that no science is at liberty to assume that its objects and constructions correspond to reality (or, rather, to prior agreements about reality). But precisely such an assumption is entailed covertly if we take the same liberties with ordinary language on behalf of linguistics as scientists in other fields are privileged to do on behalf of their sciences. The unique problem for linguistics is the unavailability of a metalanguage for the subject matter of linguistics that is separate from and prior to that subject matter. This is of course because the subject matter of linguistics is language itself, and language as a long-observed fact does contain one of its own metalanguages. Harris developed methods for describing language that do not depend upon a metalanguage imagined to be prior to and external to language itself.

5. It may be that Harris' commitment to formal methodology was initially for the sake of the "detached pattern" in language without full apprehension of the eventual fruit of this long line of research. Or it may simply be that he played his cards close to the vest, reluctant to indulge in speculation about theoretical ramifications until results were more firmly in hand. (See the discussion below of explanatory principles external to language, and how easily a plausible "way of talking" can seduce one into prematurely abandoning a line of research leading to a deeper explanation — a peril which he seems to have appreciated at an early stage (e.g. 1941:707). Harris' conception of the proper conduct of scientific inquiry is another theme calling for research and discussion.) But even if the chronology of Harris' motivations should remain indeterminate for the historiographer, nonetheless because of the consistency of those motivations and the integrity with which he enacted them throughout his long career, his work is enormously easier to understand with the benefit of hindsight, knowing now where he got to with it. And with hindsight, the numerous misunderstandings and misconstruals of Harris' work stand in vivid relief.

The dilemma posed by the absence of a language-independent, external metalanguage was the motivation, evidently, of Bloomfield's well-known rejection of mentalistic terminology in favor of studying the correlation of linguistic form with the meanings that such terminology was intended to identify, as reflected for example in (1939 loc. cit.): "It is the belief of the present writer that the scientific description of the universe […] requires none of the mentalistic terms, because the gaps which these terms are intended to bridge exist only so long as language is left out of account." See also Wells (1962:708):

6. A grammar stated in the metalanguage vocabulary and syntax of some other natural language (say, a grammar of Achumawi written in English) could as well be written in the same language, given suitable vocabulary (even borrowed if need be). The vocabulary and syntax required for a least metalanguage are quite limited (Harris 1991, Chap. 10; see also 1991:31-32).

Tabulations and mathematical or other symbolic formulations are routinely "spelled out" in oral presentations of any science. Even mathematics depends upon the "background vernacular" of language, as noted by Borel (1928:160), quoted in Ryckman (1986:289-90). And as Borel observes, the crucial role of language in all this is usually unnoticed.

7. As did Bloomfield, for instance, where he stated (1927:142) "In the study of linguistic forms, therefore, I should not appeal, as Jespersen sometimes does, to meaning as if it were separable from form", or further:

The notion of the autonomy of linguistic form from meaning is a development in Generativist theory, not traceable to Harris, much less to Bloomfield and Sapir.

8. On this point, Henry Hoenigswald (p.c.) observes: "Synonymy may not be the best example because of its degeneracy, but do synonyms have near-identical privileges of occurrences because they have near-identical meanings, or are items with near-identical privileges said to have near-identical meanings, and why would it make a difference? The hunt for discourses—words and sentences only insofar as they happen to function as discourses—that accommodate one item but not the other is, in fact, something like a pair test. Harris was creatively generalizing from Bloomfield's notion of `grammatical meaning'—an area where the unity of distributional description (in the grammar) and translating or paraphrasing (in the dictionary; itself a form of distributional statement) is traditional when it comes to things like `particles'."

9. Linguistic information is "in" utterances, socially presented in a way that is simply not available for the memories and imaginings that speakers and hearers may idiosyncratically associate with utterances, despite our assurances to one another that we have reached this agreement or that over our respective perceptions.

10. Some earlier representative quotations show that this was a view long and consistently held:

A different perspective, from anthropology and animal ethology, is brought by Gregory Bateson:

See now Harris' discussion of animal communication at (1991:382-3).

11. In the indicated place, Chap. 2 of his (1968), Harris notes "certain apparently universal and essential properties of language, which are observable without any mathematical analysis, and which are such as to make possible a mathematical treatment" (1968:6). These are that language elements are discrete, socially preset in speaker and hearer, and arbitrary, that combinations of these elements are linear and denumerable, that not all combinations constitute a discourse, that operations are contiguous, that the metalanguage is in the language, and that language changes. Each of these properties has metatheoretical consequences for linguistics.

12. Harris makes the same point in earlier writings, for example:

The application of this principle to a critique of alternative theories as a "test of relevance" (1988:57, quoted below) might be a good topic for further research. See also (1991:33-36), quoted in Section 8, below.

13. This methodological problem is usually misconstrued as an exclusion of "mentalist" concepts of meaning on empiricist grounds. Harris was neither a behaviorist nor a logical positivist. His philosophical affinities lay more with the naturalism of Dewey. (On Harris' naturalism, see Ryckman 1986:288-309.)

14. Bloomfield's commitment to social psychology and to the fundamentally social character of language did not end when he exchanged Wundt (in 1914)