BRUCE E. NEVIN http://www.cpmc.columbia.edu/zellig/NevinContrasts.htm
Bolt Beranek & Newman/University of Pennsylvania

Abstract

When we approach an unknown language, how do we know which utterances are the same and which are different? Leonard Bloomfield said that we can only tell by knowing their meanings.

One way was to try to define contrast without reference to meaning, by analyzing the distributional relations among observed phonetic features in utterances. To many American linguists following Bloomfield, this seemed to be the only scientific approach. After all, if you are concerned with a relation of contrast, you must first identify, phonetically, the things that are contrasted. On this view, the "gross acoustic features" are the observational primitives, and the task is to define phonemic contrast by analyzing the distribution of these phonetic primitives relative to one another in a corpus.

Bernard Bloch exemplified this approach. "Contrast between sounds can be defined, I think, on the basis of distribution alone, without the customary appeal to meaning" (1953:59[224]). Since "the facts of pronunciation [are] the only data relevant to phonemic analysis" (Bloch 1941:95), phonemes had to be closely identified with their phonetic detail. In his carefully worked-out postulates, Bloch based the segmentation of utterances on articulatory movements (Bloch 1948, postulates 11-16), and required that all the members of a given phoneme must have some characteristic phonetic feature in common: "The class of all segments … containing a given feature is a phoneme. […] The feature common to all members of a given phoneme is the characteristic of the phoneme" (Bloch 1948, definitions 53.2, 53.12). This constraint appeared necessary in order to determine contrast. Without it, he proposed, one could not know what was in contrast with what.

It turned out not to be so easy to define contrast in this way. Problems arose, for example, when the membership of a given phonetic segment in one phoneme or another was indeterminate (due to neutralization) or ambivalent (due to overlapping). Bloch’s (1941) follow-up brought these difficulties into focus, presenting a well-defined target for the well-known attack by Chomsky and Halle on "taxonomic phonemics" (Halle (1959), Chomsky (1964), Chomsky and Halle (1965)). However, it is important to realize that Chomsky agrees with Bloch et al. insofar as he "assume[s] that each utterance of any language can be uniquely represented as a sequence of phones, each of which can be regarded as an abbreviation for a set of features" (1964:78[407]), and that this segmentation of utterances into phones according to a matrix of universal features is the starting point for all work in phonology, whether "taxonomic" or Generative. The same assumption is found in related writings by Jakobson and Halle.

But the facts of pronunciation are not the only data relevant to phonemic analysis, nor even the uniquely fundamental data. During the same period, Zellig Harris proposed a more direct way to determine contrast without unsupported assumptions. Rather than try to define contrast, he observes it. If you want to know which utterances are the same and which are different, ask—a native speaker can reliably tell you. Furthermore, that is the only way to find out.

Harris’s aim is to disclose the correlation of linguistic form with meaning. The contrasts between utterances are the irreducible least elements of this correlation. To specify these fundamental elements of linguistics, the contrasts, Harris used a criterion of differential meaning in various substitution tests. The most precisely controlled of these substitution tests is the pair test. These tests distinguish contrast from repetition. Furthermore, and crucially, it is the substitution tests, not phonetic theory, that determine a "linguistically relevant" segmentation of utterances—linguistically relevant because the segments represent the contrasts and locate them relative to one another within an utterance. A purely phonetic basis for segmentation is not linguistically relevant because it says nothing about the correlation of form with meaning. Because the segmentation is a segmentation of (records of) utterances, phonetic detail is associated with each segment. However, the facts of pronunciation do not determine the segmentation. And because vocabulary and utterances are unique to each language, this identification of the contrasts and of elements to represent them must be done separately for each language.

This makes a fundamental change in phonemics. It means that even the most preliminary linguistically relevant segmentation, even though it may display quite a bit of redundancy, is nonetheless a representation of the phonemic contrasts of the language. The purpose of subsequent phonological analysis is no longer to define contrast (which is already given), nor to define the phonemes, but rather to refine the phonemes so that the grammar may be stated more simply or more usefully.

Harris begins with an arbitrary segmentation — "arbitrary" because it is not necessarily relevant to the distinctions between utterances which we wish to represent. Purely phonetic considerations are arbitrary in this sense, that is, they cannot ensure the relevance of the segmentation, because the contrasts are socially determined (albeit within universal physical and perceptual constraints) and because the articulations and sounds of speech, such as are studied in phonetics, are continuous, and not discrete. The segmentation is accomplished in the course of (and by means of) the substitutions just mentioned:

2. The Four "Taxonomic" Conditions

A number of Chomsky’s (1964) arguments hinge on the claim that socalled "taxonomic phonemics" adheres to four principles or conditions: linearity, invariance, biuniqueness, and a strong form of biuniqueness that he calls local determinacy. We will consider each of these in turn.

2.1 Linearity

There is a simple formulation of the linearity condition in Chomsky (1957a:346-347):

Harris violates the linearity condition whenever doing so is "convenient" for obtaining a simpler or more perspicacious grammar. Such violations are found, for example, in his treatments of simultaneous components (starting with intonation contours and other suprasegmental elements), and in his analysis of partial overlapping, of which we will see more later. In cases of vowel or consonant harmony, for example, a phonetic feature that is associated with a segment written at the end of a stretch of the utterance is spread by rule over the preceding segments of that stretch. The linearity condition does not apply to Harrisian phonemics.

2.2 Invariance

Invariance is merely a restatement of the requirement, made explicit by Bloch (in his Postulates, 1948, loc. cit.), that all members of a phoneme have some characteristic phonetic feature or features in common. Chomsky (1964:79[408]) says that the invariance condition

Harris says that this formulation is intended to be equivalent to the distinctive feature theory of Trubetzkoy, Jakobson, and their followers (Harris 1951: 125fn4, 146-149), with the crucial proviso that the elements must be defined relative to one another, as representations of contrast, rather than according to any absolute scale. This point is articulated clearly in Harris’s review of Trubetzkoy (Harris 1941:346), which to my knowledge is never cited in Chomsky’s writings:

In his discussion of biuniqueness, Chomsky (1964:82) appears to attribute to Harris and to Jakobson the weaker "relative" form of invariance that admits partial overlapping. This does not mean that Harris’s phonology was bound by the invariance condition, as taxonomic phonemics is said to be. Chomsky says (1964:79) "The invariance condition has no clear meaning unless the linearity condition is also met; I will assume, then, that it is inapplicable when linearity is violated." It follows that for Harris’s phonology invariance is not required (since linearity is not), and indeed is possible only when he happens to conform also to the linearity condition. Harris sees both conditions as useful when they enhance the integrity of the grammar as a whole, so long as the one-one relation of phonemes to contrasts is preserved. We will return to this in the section on his criteria for grouping segments (a part of the discussion of complementary distribution, below).

To the extent that the distinctive features are phonetically specified, the notion of contrast in generative grammar entails the weak form of invariance. The distinctive features are sometimes spoken of as though they were the contrasts, e.g. +/-voice is "the voicing contrast," etc. This step of reification is seductive, but unwarranted. The feature [+/-voice] is a representation of a contrast in most, perhaps all, languages. However, that contrast might be represented instead by [+/-delayed VOT] or by some other phonetic parameter. The point is that neither representation is the contrast itself. Generative phonology proposes that a set of phonetic descriptors suffices to identify the contrasts of any language, universally. However, the features of this "universal alphabet" are still labels or representations of the contrasts, not the contrasts themselves. Indeed, they can only be thought of as the contrasts themselves on a presumption of phonetic invariance. If this supposition is correct, then Chomsky’s use of the term "contrast" presumes weak invariance with respect to universal phonetic parameters such as [+/-voice], and identifies those parameters with the contrasts themselves.

Undoubtedly for Harris (as for any linguist), phonetic invariance applies to the initial segmentation of utterances. Once an utterance is segmented according to its differences with other, contrasting utterances, the resulting segments can be recognized in other utterances by their phonetic attributes, and recorded in the same way as for the first utterance. (These are not the "phones" of a phonetic transcription, because only those differences that make a difference between words are recorded.) And after segments with restricted distribution have been grouped into a set whose combined distribution is relatively unrestricted (a phoneme), a given phoneme alternant (allophone) is phonetically invariant whenever it occurs, with respect to those phonetic features that are linguistically relevant.

In summary, clearly the invariance condition does not apply to Harrisian phonemics in the way that it is used to characterize "taxonomic" phonemics.

2.3 Biuniqueness

Harris says that a number of criteria are "more powerful" than invariance (1951:65), and foremost among them is biuniqueness. The reason is obvious: the phonemes are a representation of the fundamental data of linguistics, the contrasts. Whatever is done to refine or redefine the phonemes, it must remain possible to identify from them the contrasts between utterances.

A conflict between biuniqueness and invariance is the crux of phonemic overlapping. Phonemic overlapping is a problem if you believe that invariance and linearity are preconditions for defining contrast. For Bloch this was an irresolvable problem. If he relaxed the requirement for biuniqueness (a relation between phonemes and phones), he was unable to recover the phones from the phonemes. If he relaxed the requirement for invariance he could recover the phones, but he no longer had a principalled basis for defining the phonemes in the first place.

Chomsky assumes Bloch’s sense of biuniqueness. He says (1964:80[408]) that biuniqueness

Harris identifies his initial elements not by consulting phonetic theory or an inventory of universal features, but by consulting the linguistic intuitions of native speakers in the substitution tests (including the pair test) that first establish the contrasts, and then establish a linguistically relevant segmentation of utterances. The segmentation defines the phonemes of a language by isolating the contrasts, locating them relative to one another, and identifying them with phonetic properties found at those places in the segmented utterances. Phonetic properties are associated with each segment and with each phoneme because, after all, this is a segmentation of recorded utterances.

Biuniqueness is not achieved by subsequent procedures, it is entailed by this method of identifying the contrasts. There is intrinsically a one-one correspondence between these segments and the contrasts that they represent. Each distinct utterance is represented by a different sequence of segments, and each different sequence of segments represents a distinct utterance. The "biuniqueness condition" is therefore not a formal condition holding between representations at two different levels of description, a phonetic level and a phonemic level. Rather, it is a methodological requirement that one not lose or obscure the fundamental observational data for a science of language, namely, judgments by native speakers as to which utterances are repetitions and which are not.

The "sounds" of Harris (1944a) are equivalent to the segments that Harris (e.g. 1951:65fn14, 80.0, 85fn16) groups into a given phoneme while preserving a one-one correlation of phonemes to segments. The segments (with their associated "sounds") were set up at an earlier stage of analysis to be linguistically relevant by virtue of their one-one correspondence to the contrasts (either directly or by way of the segments used in still earlier stages of analysis). In other words, the biunique relation of representations to contrasts is transitive through successive reformulations of the representations. Any biunique correspondence of phonemes with "sounds" preserves the previously established biunique relation of those linguistically relevant "sounds" with the contrasts. Harris (1951) occasionally speaks of a one-one relation between phonemes and segments. This is a telegraphic usage, to avoid awkward repetition of the definition of relationship (1951:34-35) quoted earlier, where, after all, he had said "for convenience, we will set up as our elements not the distinctions, but classes of segments so defined that the classes differ from each other by all the phonemic distinctions and by these only."

When Harris describes the grouping of segments into a phoneme, it might appear that one of the requirements involves creating a biunique relation between the phoneme and its member segments, as though he were talking of a biunique relation of phonemes to phones. In fact, the requirement is not to create but to preserve the biunique relation between the phoneme and the contrasts that its member segments had. We will return to this under the topic of complementary distribution.

The biuniqueness condition, as it is used to characterize "taxonomic" phonemics, does not apply to Harrisian phonemics. In "taxonomic" phonemics, biuniqueness is a one-one correspondence of phones to phonemes; for Harris, it is the contrasts that have a one-one correspondence to the phonemic symbols (and associated phonetic properties) that locate the contrasts relative to one another in utterances.

2.4 Local Determinacy

Local determinacy is a context-free form of local determinacy: "neighboring sounds" cannot be used in defining a phoneme and partial overlapping is disallowed as well as complete overlapping. Chomsky says (1964:81-82[409]) that he is attempting to restate in clear and explicit form a widely held but inchoate view of what "taxonomic" linguists thought was necessary to maintain linguistics on a scientific footing. Local determinacy is a one-one correspondence:

It is also possible to define local determinacy in negative terms as biuniqueness with no "mixing of levels":

Chomsky’s reason for defining local determinacy as he did is that separation of levels is a methodological condition, and local determinacy is a substantive (or formal) condition. What does this mean? At the outset of Chomsky (1964), he set up a dichotomy between acquisition models and perceptual models. An acquisition model determines the correct choice of a phonemic system in accord with methodological conditions. A perceptual model relates a phonemic system to speech sounds in accord with substantive or formal conditions. Separation of levels is a methodological condition for learning the phonemic system, or for a linguistic discovery procedure. Local determinacy is a formal or substantive condition for recognizing phonemes in the phonetic continua of speech once the phonemic system has been determined (or learned). Chomsky argues that the former can shed no light on the latter. Setting aside this rather peculiar sense of "methodological"—he has told us that he is "not concerned with … methods of investigation (analytic procedures)" (1964:7)—we may still wonder at this claim that speech recognition and speech production are not subject to constraints of the same sort.

However, this interesting and now forgotten bit of polemical scaffolding has no bearing on our present discussion. Harris is very much concerned with the methodology of linguistics in the sense of methods of investigation, not language acquisition. The only indication of local determinacy in Harris’s work is as a practical matter: one should extend the environments that one tests no farther than is sufficient. Indeed, a rigid condition of local determinacy would contravene the characteristic "bootstrapping" approach to linguistic analysis that Harris followed in all his work, in which a first approximation is later refined by criteria that could not be applied or could not be defined at the earlier stage, in which tentative guesses as to the results of later stages of analysis (e.g. morphology) are used as guides at an earlier stage, subject to correction when the later work is carried out more fully, and in which earlier results are subject to re-evaluation in the light of results at a later stage, always with an eye to the overall simplicity of the grammar. This is what Harris means when he says that operations must be "carried out for all the elements simultaneously" without any "arbitrary point of departure" (1951:7). Thus, subsequent rephonemicization could take into account the results of later stages of analysis, including phonemic juncture and the boundaries of morphemes and words.

The local determinacy condition does not constrain Harrisian phonemics. Harris did not have qualms about the scientific status of invoking such entities as junctures and boundaries, and even morphophonemic alternations, both in preliminary guesses and also later for the sake of rephonemicization, because all his results refer back to the data of contrast (biuniqueness, properly understood). Given this touchstone of validity and scientific rigor, Harris was free to "bootstrap" his description by later refinements of early approximations and guesses without loss of methodological rigor. This freedom is unavailable to linguists for whom contrast is not an observational primitive, but rather something to be defined by distributional analysis of phonetic primitives. We will turn next, then, to the procedures of distributional analysis.

2.5 Summary

These four conditions are obligatory if one has only "the data of pronunciation" to work with—without linearity, invariance, biuniqueness (in the phone-phoneme sense), and local determinacy, one would be at a loss to say what might be in contrast with what. They are optional for Harris because his fundamental data are native speakers’ intuitions of contrast, identified and located in elements of a representation by the substitution tests, including the pair test. The question is not what contrasts with what, but rather what can be used to represent the contrasts that are observed (this utterance is not a repetition of that one), locate them relative to one another, and associate them with phonetic features of speech. In Harris’s (1991) view, the same question faces the child learning its first language.

3. Complementary Distribution

Chomsky says that complementary distribution is "the central concept of taxonomic phonemics as developed, for example, by Jones, Troubetzkoy, Harris, and Bloch," and attacks it as (1964:91[414]):

But we already know that Chomsky’s concept of biuniqueness refers to a correspondence of phones to phonemes, rather than a correspondence of contrasts to segments of contrasting utterances, and we have seen that local determinacy refers to qualms about mixing of levels that did not concern Harris.

From this unpromising beginning, we will step through Chomsky’s argument against complementary distribution.

3.1 Tentative Phonemes

Chomsky describes Harris’s application of complementary distribution as follows (1964:92[414.3], italics in original), claiming to summarize Harris (1951: Chap. 7):

However, this shift of perspective has a serious consequence for his reading of Harris. The comparisons of D(x) and D(y) are recursive, not merely successive, pairwise comparisons of phones. Chomsky describes a linear sequence. One imagines a sort of distributional analysis component that outputs a set of tentative phonemic systems as alternative candidates. These candidate phonemic systems are input to a test component, where they are subjected first to a test for exhaustiveness, then to "additional criteria of symmetry". The output from the test component is the winner, "the phonemic system (or systems)". Chomsky objects (1964:92-93[414-415]) that in some cases "the class of ‘tentative phonemic systems’ … will not include the optimal biunique system as a member, so that no supplementary criteria will suffice to select it from this class."

But Harris describes a massively parallel recursive process, not a linear sequence. The criteria are not applied to a set of alternative phonemic systems that are the outputs of distributional analysis, they are applied at every step of merging a segment with restricted distribution into a partially defined phoneme, whose distribution it complements and therefore adds to, resulting in a new phoneme whose distribution is less restricted.

The comparison of [x] and /y/ is one step of this recursive process, where [x] is a segment and /y/ is a phoneme-in-the-making. If the environments of segment [x] are complementary to those of phoneme /y/, and if the criteria indicate that an [x]- /y/ grouping is superior to other possible combinations (which are being tried in parallel), then [x] is merged into /y/. That is, the environments of /y/ now include those in which segment [x] occurred. It is to the now even less restricted tentative phoneme /y/ that other segments [z], etc. are subsequently compared.

The result in the end is of course just as Chomsky describes (on one reading of Chomsky 1964:92, anyway): each pair of the member segments of any given phoneme are complementary to each other, since complementarity is transitive over this succession of comparisons. However, the critical issue concerns not the end result but the recursive process of attaining it. At each step of the process, the environments of a tentative phoneme /y/ are augmented by those of a new segment [x]. Then, before proceeding further, all the environments that formerly contained the now-included segment [x] must be restated in terms of the redefined phoneme /y/ ("Adjusting Environments in the Course of Phonemicization", Harris (1951:62)). Chomsky reframes this rather obvious housekeeping step as an ad hoc procedure brought in just to save taxonomic phonemics from its flaws. In the light of the above discussion, the error can be clearly seen, for example, in (Chomsky 1964:87):

3.2 Criteria for Grouping Segments

The "additional criteria of symmetry" referred to in the above quotation of Chomsky (1964:92[414.3]) are not criteria for evaluating alternative "tentative phonemic systems" produced by analysis of the complementary distribution of phones relative to one another. They are criteria for making each choice in the course of grouping segments, one by one, into partially defined phonemes (Harris 1951:62-63). Another quotation affords a convenient context for reviewing these criteria. Chomsky argues (1964:77[407]) that the "phonetic substance" of phonemes cannot be supplanted by distributional or other criteria for grouping segments into phonemes:

When we undertake the work leading to a more efficient representation for the contrasts, "in most cases there will be more than one way of grouping segments into phonemes […] It is therefore necessary to agree on certain criteria which will determine which of the eligible segments go together into a phoneme" (Harris 1951:63). Harris’s concern at this stage of analysis is to guide the process of grouping segments so that a simpler and more useful description is possible. The criteria that he applies to this end may be thought of as evaluation metrics, with the caveat that we have just noted in the preceding section: one should not think of some sort of ‘distributional analysis component’ churning out a set of candidate tentative phonemicizations, followed by an ‘evaluation component’ that applies the criteria and selects a winner as the phonemic system.

Most of the evaluation criteria by which Harris proposes to guide distributional analysis (section 7.3) to a more efficient representation for the phonemic distinctions (contrasts) are stated in terms of symmetry in the representation of sounds (section 7.42) and symmetry of environment (section 7.43), but in each case Harris clearly and unequivocally states that the motivation for the given criterion is not merely symmetry for its own sake but rather the simplifications that symmetry makes possible in the descriptive statements (rules) of the grammar. One metacriterion (condition for applying the other criteria) concerning the simplicity of the description is as follows:

The three criteria that hinge on phonetic considerations ("symmetry in the representation of sounds") are as follows:

1. "Identity of representation among segments," that is, their having phonetic features in common (section 7.421). This is "convenient" because it makes for a simpler phonological description to be able to "speak of the phoneme as representing this common feature, rather than as being a class of segments. Relations between phonemes would then represent relations between sound features."
1. "Identity of inter-segmental relation among phonemes," that is, having the same phonetic difference between parallel pairs of allophones (section 7.422). "It is also convenient to have the relation among segment definitions within one phoneme identical with the relation in other phonemes." This makes for great simplification in the phonological description, because a single statement may account for the parallel allophony of multiple phonemes.
1. Choosing phonetic features that apply across the "complete phoneme stock" (section 7.423). "If the objective is a minimal stock of phonemes, the definition of each of which is to be as simple as our other criteria permit, it follows that the selection of the common features should be governed by the generality of these features and differences among all the segmental elements of the language." This is "convenient" because it broadens the set of phonemes to which a statement (rule) affecting a given feature can potentially be made to apply, simplifying the grammar. "We can discover which groupings [of segments] will yield the most simply defined phonemes by testing the differentiation, upon which we propose to assign particular segments, throughout all the segments."

When one reads what Harris was actually saying, Chomsky’s interpretation of it as "suggesting that a non-phonetic principle can replace reliance on absolute phonetic properties" in this way seems at least far-fetched. Nevertheless, it is also clear that for Harris the phonetic properties of segments are subordinate to the fundamental relationship of contrast.

There is another criterion that Chomsky neglects to mention, namely, the criterion of morphemic identity (appendix to 7.4) with the obvious benefit of simplifying morphophonemics.

Clearly, Chomsky’s claim that "the correct analysis is simpler only if we utilize the familiar phonetic properties for phonetic specification" is false, and betrays either a misunderstanding or a misrepresentation of Harris.

3.3 Complementarity and Biuniqueness

Chomsky characterizes complementary distribution (1964:91[414]) as a procedure whose "goal is to provide the minimally redundant representation meeting the conditions of biuniqueness and local determinacy." Minimal redundancy is indeed an important aim of Harris’s procedures, in order that the correlation of form with meaning should be as direct and transparent as possible. However, a set of elements that correspond one-one to contrasts between utterances is not necessarily minimally redundant. Many potential systems of representation might bear a biunique relation to the contrasts between utterances, including an initial segmentation determined (non-uniquely) by Harris’s substitution tests.

All such representations are ‘phonemic’ in the essential sense that the contrasts are phonemic distinctions. Most of these possible systems of representing the contrasts have more elements than are needed, and the elements are more restricted than necessary in their combinations with one another.

Harris himself said (1951:62-63) that complementary distribution alone is not a sufficient criterion to guide the recursive process of grouping segments into phonemes to a preferred or optimal representation, and as we have seen he deployed a range of criteria for grouping segments so as to represent the phonemic contrasts more efficiently, that is, so as to simplify the grammar ("in most cases there will be more than one way of grouping segments into phonemes […] It is therefore necessary to agree on certain criteria which will determine which of the eligible segments go together into a phoneme" (1951:63)). It is difficult, therefore, to motivate Chomsky’s demand that complementary distribution, reduced to a mechanical discovery procedure, should alone "provide the minimally redundant representation".

This is nonetheless the only possible construal to put on Chomsky’s ensuing discussion of complementary distribution, as he turns to an example of phonemic overlapping due to Bloch. In the dialect that Bloch describes, alveolar flap [D] occurs intervocalically after stress in e.g. “Betty”, and after ? in e.g. “throw”. Chomsky says (1964:92[414]):

The relation between contrasts and representations in this example may be considered as follows:

Word Contrasts:	"Betty"	"berry"	"throw"
Representation #1:	beDiy	[beriy]	[?Dow
Representation #2:	betiy	[beriy]	[?row]

The relation between contrasts and representations in this example may be considered as follows:

Word Contrasts:	"Betty"	"berry"	"throw"
Representation #1:	beDiy	beriy	?Dow
Representation #2:	betiy	beriy	?row

In Representation #1, the distribution of the segment D is restricted to two environments, intervocalic and following ?. At the same time, the environment of the element t is restricted because it lacks the intervocalic environment, and the environment of the element r is restricted because it lacks the environment after ?. In Representation #2, all these restrictions are removed by adding to /t/ the intervocalic environment where D occurs and by adding to /r/ the environment after ? where D occurs. (We will discuss the formal basis for dividing the occurrences of D presently.) The segments in beDiy, beriy, ?Dow constitute a less efficient representation of the contrasts between words in English (including these three words) than do the segments in betiy, beriy, ?row, but both representations preserve a biunique relation to the contrasts. Chomsky is either misrepresenting Harris, or he has not understood him.

3.4 Neither Necessary Nor Sufficient

Chomsky continues (1964:92-93[414-415]):

Crucially for this example, Chomsky ignores Harris’s description of dividing the distributional range of a segment. We divide the [D] segment into two distributionally defined (but phonetically identical) elements, and assign the one that occurs after [?] to /r/ and the one that occurs intervocalically after stress to /t/. This procedure is a further extension of distributional analysis, yielding a new system that had previously not been considered. Harris describes this procedure and its justification as follows (1951:91-92):

However, the thrust of Chomsky’s argument against complementary distribution is that it is not sufficient to determine contrast. This is explicit in his next example (415.1):

There is a perhaps more obvious problem with Chomsky’s example of a complementary grouping of [k] and [a] into /K/, however, and that is that it would never be pursued very far, even by a mechanical discovery procedure (if there were such a thing), because it does not generalize. A tentative phonemicization that groups [k] with [a] exploits the complementarity of consonants with vowels. But having exploited it for the [k]-[a] pair, that complementarity is no longer available for all C-V pairs as a class. Major distributional regularities would be lost in favor of a smaller and more restricted grouping that adds to the complexity of the description. Descriptive statements (rules) could no longer apply to the class of vowels, or of consonants, or of stops, and so on. Thus, it is scarcely surprising that this socalled "problem has received little attention" (Chomsky 1964:93[415]).

Every step of this argument having fallen apart, Chomsky’s conclusion falls too (1964:93[415]):

Chomsky next takes up Harris’s discussion (1951: 62, section 7.31) of the need to redefine environments during the course of distributional analysis. When a segment being tested against environments is merged with others in a phoneme, before one can proceed with further analysis the environments for further testing must be rewritten, so that this segment is replaced by the phoneme everywhere that it occurs. We briefly discussed this recursive property of distributional analysis earlier. In a footnote (1951:62fn10, substituting small caps for subscript [û]), Harris points out the consequences of overlooking what is after all an obvious requisite for carrying out the work of linguistic analysis in a systematic and logical way:

3.6 Ou Tout Se Tient

There is a broader methodological issue lurking behind this discussion, concerning the handling of data in relation to emerging results. Chomsky’s attack on distributional analysis continues (ibid.):

Having diminished a broadly applied housekeeping principle to the status of an ad hoc "procedure" aimed at rescuing distributional analysis from awkward counterexamples like the grouping of [t] and [k] into /T/, Chomsky argues that such a procedure violates a global requirement that Harris must retain in order to avoid use of rule ordering. Continuing (ibid.):

Harris continues (1951:7, emphasis added):

Chomsky’s rather curious suggestion that Harris wishes to avoid rule ordering is in his next sentence at the cited place (1964:94[415]):

3.7 Chomsky’s "Condition C"

Chomsky closes his survey of "taxonomic phonemics" in general and of complementary distribution in particular with the following generalization, labelled "condition C" (1964:95[416]):

Chomsky’s Condition C states the case in terms of phone sequences that contrast. This amounts to substituting Chomsky’s "universal phonetic alphabet" in place of Harris’s segmentation based upon substitution preserving contrast/repetition. Harris’s initial segments are defined relative to one another, but Chomsky assumes the initial segments are "phones," elements defined in phonetic terms. Harris of course uses terms of phonetic theory as descriptors of phonetic properties associated with the segments, but the phonetic properties do not determine the segmentation or the linguistic relevance of the segments. Harris’s segmental representations of the contrasting utterances always necessarily differ, no matter whether they count in Chomsky’s reckoning as "phone sequences," as "tentative phonemes," or as phonemic representations. They always differ because Harris’s representations of utterances are representations of contrasts between utterances, and therefore necessarily have a biunique correspondence to the contrasts between utterances. Chomsky’s Condition C is met from the outset.

The satisfaction of Condition C is preserved under distributional redefinition of the phonemic elements, not created by it. The procedures of distributional analysis ensure that each subsequent redefinition, refinement, and rephonemicization of the representation preserves the biunique correspondence that the prior one held with respect to the contrasts. The relation of Biuniqueness is transitive all the way back to the initial representation, such that each new representation of utterances X and Y a fortiori has a one-one correspondence to the primitive contrasts between utterances X and Y. As a consequence, "if phone sequences X and Y contrast", then the representations of X necessarily differ from the representations of Y under all these redefinitions, so long as Harris’s procedures are followed. The criteria of simplicity and symmetry (in the several senses of 7.421, 7.422, 7.423, and 7.43) help to determine one of the solutions that best meet the overarching criterion of simplicity or efficiency.

4. The Argument From Rule Ordering

Surrounding the discussion of "taxonomic phonemics" is a presentation of Halle’s argument (Chomsky, 1964:88[412-413]) "that it is generally impossible to provide a level of representation meeting the biuniqueness condition without destroying the generality of rules, when the sound system has an assymetry." To demonstrate this, derivations of different forms are laid out in parallel, showing a segmental representation after each application of rules. (In the following adaptation of Halle’s example, diacritics are omitted.)

Iin Russian, voicing is contrastive for stops but not for the affricate. (This is the "assymetry" in Chomsky’s statement.) A rule that voices obstruents before voiced obstruents is morphophonemic for stops, but allophonic for the affricate. The morphophonemic and allophonic rules are identical in form, they differ only in their scope. Because there is no voicing contrast for affricates, the voicing rule affecting them is allophonic, whereas the rule affecting stops is morphophonemic. For each segmental representation that we might propose as the phonemic "level of representation", it is shown that before that level some rule applies as a morphophonemic rule to some forms, and that after that level an identical allophonic rule applies to other forms.

This difficulty dissolves as soon as you realize that the distinctive features are the representations of contrasts. Harris had observed that his simultaneous components in general (1944a:205, 1951:133.3) and distinctive features in particular (unit-length components defined for the whole stock of phonemic contrasts of a language, 1951:147.2) may supplant the segmental phonemes.

Chomsky proposes to eliminate the phonemic "level" of representation:

i.	physical phonetics
ii.	systematic phonetics	(distinctive features)
iii.	"taxonomic phonemics"	(segments, eliminated)
iv.	systematic phonemics	(morphophonemic representation)

In fact, the representation of phonemic contrasts is simply shifted from "taxonomic" segmental phonemes to distinctive features. What Chomsky calls "systematic phonetics", using distinctive features, is no less phonemic than the segmental representation that it supplants. The effect of the actual changes in representation of sound systems is as follows:

5. Consequences for Grammar and for a Theory of Language

Halle (1954:335) and Chomsky (1957:234[343]) refer to the pair test as a fundamental starting place for identification of repetitions, but they do not recognize that the substitution tests (including the pair test as a special case) are the basis for segmentation of utterances. Nor are the ramifications of contrast being given in advance developed in Generative phonology. Instead, the notion of contrast is presented just as in the work of Bloch and other "Neo-Bloomfieldians": a function of phonetic differences between physically defined segments, as given by an antecedent study of phonetics. Chomsky (1964) does not say how the segmentation is done. Halle (1954) seems to suggest that to define an inventory of segments we should (1) use the pair test to partition the set of utterances into repetitions and non-repetitions, (2) create an exhaustive roster of minimal pairs, (3) project matrices of universally distinctive features onto the minimal pairs.

The shift from a segmental representation to a distinctive feature representation changed the character of descriptive statements of phonology and in particular the character of rules deriving phonetic descriptions from morphophonemic representations. It is true that Bloomfield (1933) had conceived of the phoneme as a bundle of distinctive features, but it appears that he thought of the bundle (the phoneme) as the fundamental thing, an "indivisible unit" (1933:79), and the features as a part of their descriptive analysis. In the proposals of Jakobson, Fant, Halle, and here Chomsky, the converse is true: the distinctive features were seen as fundamental, and segmental (alphabetic) representations were seen as convenient notational abbreviations. Indeed, Chomsky & Halle (1965:472) argue that a segmental representation may be dispensed with entirely:

                                        t’           e                     t           o                     k’^w           e                     k^w          o
laryngeal
          [spread]                    +                              +                              +                              +
          [constricted]
          [voice]                              +                              +                              +                              +
[nasal]
oral cavity
          aperture                                                                                                                                  A₀
                    [open]                              +                              +                              +                              +          A_f
                    [open]                              +                              +                              +                              +          A_max
                    [open]                              +                              +                              +                              +          mid
                    [open]                                                                                                                                     low
          [labial]                                                            +                    +                              +          +
          [coronal]
                    [anterior]                    +                              +
                    [distributed]                    (+)          +                                                  (+)          +
          [dorsal]                                                            +                                                            +
                    [posterior]                                                                                 +                              +

As a notational abbreviation, parentheses show leftward spread of the [+distributed] feature. The corresponding matrix for the Indo-Iranian CV sequences is easy enough to construct, so that I won’t belabor the obvious point that this notation is rather less convenient for writer and reader.

Harris could define IE */e/ so that it includes a long component of palatalization that extends over a preceding consonant. This is functionally equivalent to the leftward spread of a feature, differing only in where in the representation of contrasts the information is located (the definitions of segments vs. rules that apply to feature matrices that specify segments).

The salient point is that palatalization, or the [+distributed] feature, is retained in the CV sequence IE */k^we/ > Indo-Iranian */?a/ in its entirety, (i.e. the vowel */a/ of the latter is presumably palatalized or has a palatalized onset), and it is lost from the CV sequence IE */te/ > Indo-Iranian */ta/ in its entirety. The perceived shift of the feature from the V segment IE */e/ to the C segment Indo-Iranian */?/ is a notational artifact, a decision in each case as to the location of the contrast that holds between utterances in which these segments occur.
 

IE	*/te/ *[t’e]	*/to/ *[to]	*/kwe/ [k’we]	* /kwo/ [kwo]
Indo-Iranian	*/ta/		*/?a/	*/ka/

A purely phonetic basis for segmentation is not linguistically relevant because it says nothing about the correlation of form with meaning.

Re Halle’s Russian example, it is the same rule for a difference that marks a contrast and a difference that does not, but who cares? There is still a difference between contrastive and non-contrastive forms.

In Harris’s view, neither the features nor the segments have any privileged ontological status. It is the contrasts that are "real". This broaches the peculiar ontological status of language. In the familiar Berkeleyan example, if a tree falls in a forest and no one is there to perceive it, we nonetheless assume the reality of the tree and of the event, and, insofar as sound is defined in physical terms as pressure waves in the atmosphere, there is a sound. However, if a tape recording of the Gettysburg Address is played in the forest and no one is there, there is only sound. There can be no language present—no words, no contrasts or distinctive features—in the absence of a hearer who controls perceptions of the words, contrasts, etc. of English. The contrasts may be represented by phonemic segments, phonemic components, distinctive features (unit-length phonemic components), or in some other way, but they cannot be defined absolutely, in purely physical terms. They can only be defined relatively, in terms of the contrasts between utterances that they represent. And the contrasts are not physically given, they are socially given. This is why the pair test is necessary for determining the phonemic distinctions or contrasts in a language. The means for making contrasts may be universal; the contrasts are as language-particularized as the vocabulary in which they may be located.

In Chomsky’s view, the features are real in the sense that they are built into the human brain according to inherited properties of the human genome. They are genetically given. But the features are the available means for making contrasts. The contrasts themselves must be determined for each language by the pair test.

For Harris, phonology is encapsulated, in the sense that to work on syntax and semantics it does not matter what representation is used for the phonemic contrasts, ordinary orthography will do; and clearly this is also the case in practice for Generativist syntax.

For Chomsky, feature notation is required because PSG drives phonetic content out at the bottom just as it drives semantics out at the top. In operator grammar, phonetic content associated with segments is present from the moment of word entry in the base. For Harris, For Chomsky, phonology, morphology, and syntax are interimplicated in a fragile structure of great complexity.]

[Features are useful for stating universals.

The purpose of this paper has been to to set the record straight and to explicate Harris’s insights into the nature of linguistic contrast. The consequences of reinstating these insights appropriately in the theory and practice of linguistics today go beyond what it is possible to consider here.