CHAPTER TEN

How Different Are They?

Fred W. Householder

With the exception of the paper on Pāṇini, all the participants in this conference are implicitly criticizing or, at least, expanding on the phonological theory presented in Sound Pattern of English (Chomsky and Halle 1968). It behooves us, therefore, to start by specifying as explicitly as we can the principal tenets of this theory, which cannot be found on any one page of SPE (or anywhere else, to the best of my knowledge). We can then note where and in what ways each of the conferees has chosen to differ from SPE (and it must be said that few of them differ on any really fundamental issue), and we may also point out choice axioms for future “phonologies” to reject. It must be noted that few of the “phonologies” of this volume differ substantially on fundamentals from the SPE theory, though perhaps Sanders’ Equational Phonology comes closest (except for Pāṇini, of course) to an appreciable difference; indeed for some, such as Dinnsen and Houlihan-Iverson, it is hard to say that they differ at all on essentials. If we had had a representative of Pikean or Firthian or stratificational phonology, or of New Wave or Labovian phonology, the differences would be more conspicuous.

I will try to subsume the principle axioms of SPE phonology under five main heads, listing relevant page numbers after each paragraph.

(1) The task of the linguist (as grammar writer for a particular natural language) is to duplicate in explicit notation the grammar which is implicitly present in the idealized speaker-hearer’s brain, a grammar which he acquired as a child by the use of a number of specialized innate language learning devices. The explicit notation must somehow correspond functionally to some of the innate devices. The idealization involves, among other things, (a) the selection in most cases (of words, inflections, syntactic patterns, etc.) of a single one of the known optional or dialectal or stylistic variants as the correct word, form, etc., (b) the inclusion of all grammatical competence and etymological information possessed by any speaker of the language—not excepting grammarians and etymologists, and (c) the assumption that language acquisition is instantaneous. (3, 4, 25, 322, 331, etc.)

(2) In dealing with the phonological part of his task, the linguist (i.e., phonologist) must provide the correct lexical representations, distinctive features, and phonological rules to map terminal strings of the syntax onto systematic phonetic strings. The answers to questions about the systematic phonetic structure of any language are intuitively given (i.e., segmentalization, choice of features, inventory of segments, etc.); for English in SPE, these answers closely resemble those belonging to the Smith-Trager autonomous phonemic analysis of English. The same set of features must be used throughout, though the kind of values (binary, m — u, +—, n-valued, etc.) may change. The rules must be ordered in a single sequence, with no back-tracking except perhaps for cyclical application. (5, 9, 11, 12, 14, 23, 25, 75, 296-98, etc.)

(3) The innate universals include a set of universal features, a set of rules and restrictions for combining them into segments, a set of permissible phonological rule types, the cyclic principle, a natural drive to eliminate allomorphy by means of rules, and an evaluation metric to rate alternative rules or rule sets. The phonologist must use all and only these devices used by the language-learning child, (ix, 18, 19, 43, 249, 251, 295, 296, 297, 333, 335, 356, 364, etc.)

(4) The primary function of the rules is to account in a phonetically plausible way for all the surface forms (i.e., inflexions or contextual variants) of a given lexical item and all derivatives from a given etymological root (or whatever you may wish to call it) with a single minimally specified base form. Suppletion or listing is to be resorted to only in rare and special cases, (ix, 381, 388, and passim)

(5) The linguist, whether as phonologist or syntactician or semanticist, need not concern himself with the uses or functions of language. (No mention at all in SPE; the notion “natural” appears on 335, 400, etc., and “plausible” on 401, 419, etc., but neither term is linked to function.)

As we remarked above, Pāṇini’s phonology is perhaps the most strikingly divergent from these principles; there is no reason to suppose that he would have subscribed to any of section (1) or (3) or (5), and while he certainly does use the single base (at least for grammatical morphemes, and in large part also for ordinary words, as represented in the list of roots—Dhātupāṭha—associated with the Aṣṭādhyāyī), it is considerably more abstract than SPE theory allows, and the rules that operate on it form a kind of Unrestricted Rewrite System; nor does he appear to view correctness as an empirical matter in the sense of SPE: 331. Rather he writes what we would describe as a linguist’s grammar, aiming at maximum economy with at most a kind of weak equivalence to a speaker’s internalized grammar. A good deal of structural phonology in the forties and fifties was written from the same (not necessarily hocus-pocus) point of view. Pāṇini’s methods of ordering the application of rules as here presented by Joshi-Kiparsky do not, in the main, coincide with those of SPE, though the principle of “proper inclusion” or “elsewhere” discussed here by Anderson is certainly important. It, too, was commonly used by the structuralists.

Since Anderson’s paper constitutes, in part, a history of the developments within generative phonology during the last ten years, I shall deal with it first. Much of it deals with the earlier activity of contributors to this volume and of others who could have been, but much also deals with his own activities, which seem to involve four matters which go beyond the relatively minor issue of mechanical principles of rule application (including a rejection of the cycle): (1)arguments for the necessity of a set of morphological rules distinct from the phonological rules as well as phonetic rules and “quasi-systematic relations” among lexical items; (2) arguments against the strict linear ordering of rules and in favor of principles to determine relative ordering in particular cases (both of these are minor details of paragraph (2) above); (3) the proposal of a principle of exegetic adequacy (i.e., adequacy for retrospective explanation rather than prediction); (4) a rejection of the notion of complete universal naturalness (as implemented by the marking schemes of SPE: chapter 9). All these, though touching none of the major points of the theory, seem to me to be distinct improvements on SPE.

Goldsmith’s Autosegmental Phonology has two layers: at the first level it merely accepts the SPE invitation to consider matters of tone which were completely avoided by Chomsky and Halle, but at the second it offers a potentially radical revision in the nature of both the lexical representation and the systematic phonetic one, which are conceived in SPE as single-stranded concatenations of segments each of which consists of a cluster of simultaneous features. Goldsmith at least envisages the possibility of several parallel concatenations with variable linkages (“association lines”) between segments in one and segments in another. Again no fundamental axioms of SPE are altered. The proposal suggests various testable speculations about language acquisition by children and eases the formalization of various changes, particularly assimilations, some of them also improvable by theories such as R. Cheng’s (1977), which make the sequence single but the units in it syllables, each syllable being a bundle of simultaneous features which may be arranged in various ways on the surface. Many of the same goals may be attained either by multiplying the strings vertically or by broadening the beads horizontally.

Atomic Phonology (Dinnsen) and Functionally Constrained Phonology (Houlihan-Iverson) make virtually no changes in the basic theory of SPE, nor do they expand it in the manner of Autosegmental Phonology. Instead they each attempt to provide some restriction on the nature and content of phonological rules. These are proposals for universal, in some sense. Dinnsen’s proposal has to do with the notion of complement rule-pairs: these are two rules which (essentially, though redundant features may alter things slightly) differ only in regard to the value (+ or —) of a single feature in the input (left side of environment). For all such pairs it is claimed that only one will be attested—i.e., if there are languages which drop final labials after rounded vowels, there will be none which drop labials only after unrounded vowels. The member of the pair which exists is called an atomic rule, the other member a nonatomic rule. The only test for atomicity appears to be heuristic; look through the languages of the world. (Incidentally, Charles A. Ferguson sent all the participants in the conference an invitation to use the Stanford Phonology Archive to search through 200 languages for data of this sort; so far as I know, none of them has done so yet.) Though nonatomic rules (without their mates) are said to be non-existent, all rule generalization or simplification is said to take place by the merger of an atomic rule with its mate, i.e., in the example mentioned above, a change to a rule dropping final labials after all vowels. Though Dinnsen says this applies also to generalization by Greek letter variables, it would seem that pairs which could be so generalized would involve opposite values in two places, often on both sides of the arrow; e.g., a rule that drops only labials after rounded vowels would merge with one that drops only non-labials after unrounded vowels. This seems to require a redefinition of the term “complement rule.” Another claim made in the Dinnsen article, though it does not seem to be connected with atomic or complement rules, is that all intervocalic voicing rules are allophonic, i.e., do not merge phonemes which are elsewhere distinct. But in the passage from Latin to French somewhere, intervocalic p and b, which remain distinct initially, merge, ending up as v (and f and v are also distinct initially). Of course, once they have merged, there will be no contrasting segments in the input any more.

The Houlihan-Iverson paper makes proposals very similar to Dinnsen’s, and comes no closer to being a new or distinct phonological theory. Their main principles seem to depend on a notion of markedness which is quite different from that in SPE (where the marked value of a feature may vary according to context), and which they define solely in terms of implicational universals: if there are languages with value a and without — ⍺, but none with only the — ⍺ value, then the a value is marked (presumably in all positions as well as all languages). There has always been a hedge available to phonologists when new data appear to upset an implicational universal—they can simply say that a different feature is involved. “It’s not voicing here, it’s tenseness.”

Their second main point has to do with the difference between neutralization effects and allophonic ones where they choose the rather specialized Kiparskyan definition of neutralization. In these terms certain rules are claimed to be always neutralizing (e.g., final devoicing), others always allophonic (e.g., intervocalic voicing, as in Dinnsen). It is interesting that final devoicing is said to be always neutralizing, in view of the original Halle-Lees argument against taxonomie phonemes (that a taxonomie system could not capture in one rule the universality of assimilative voicing in Russian, since for some phonemes—including [x], [ɣ]—it was only allophonic, while for others it was phonemic). Houlihan-Iverson however claim that final devoicing—including [ɣ] → [x]—is always neutralizing in Russian, citing a good taxonomie phonemicist, George Trager. Azerbaijani, however, has a final devoicing rule which affects not only stops and ordinary fricatives (for all of which a good case for phonemic voicing exists) but also the phoneme /r/, which certainly does not have contrastive voice. So here Houlihan-Iverson would require two rules, one for the other obstruents and one for /r/ (whose voiceless final allophone is, in fact, a fricative). In their Corsican example, however, where word-initial voiceless stops become voiced after preceding vowels while word-initial voiced stops become fricatives (presumably by a rule which is extrinsically ordered earlier, in the SPE manner), they have apparently altered their definition of “in the input” in Kiparsky’s definition of neutralization so that it no longer means “existing at some level of derivation before the rule applies,” since, if they stick to that, the Corsican rule must be neutralizing.

Among their implicational universals (which refer, as they say, to inventories of phonetic segments), one, at least, seems to be questionable, (7)f “The presence of mid vowels implies the presence of high vowels,” since several Amerindian languages (e.g., Apache, Mazateco, Potawatomi; see Hockett 1955:81-5; Ferguson’s computer program could surely supply more) have [o] as their highest back vowel. Perhaps the universal could be interpreted to mean “some high vowel,” not “a corresponding high vowel,” and indeed Apache, etc. do have an [i]. This definition of markedness in terms of implicational universals forces them to maintain that voicing (of obstruents) is equally marked in all positions, but it has long been accepted that voiced obstruents are very frequent in intervocalic position, while voiceless ones are especially common initially and finally, and this distribution of allophones is actually found in some languages which lack a voicing contrast. (I discussed this matter a little in Householder 1971.)

Although they do not quite claim that neutralization to a marked member never occurs, they certainly suggest that; and if fricatives are always marked relative to stops, it is interesting that final velar and palatal stops in Azerbaijani are always neutralized to the corresponding fricatives, and this cannot be allophonic for the velar fricative. Examples contrary to other Houlihan-Iverson predictions were provided at the meetings by Jonathan Kaye—one from Quechua, others from Algonquin and Montagnais.

The one paper which presents the maximum of notational innovation, but also resembles in various ways the Dinnsen and Houlihan-Iverson papers, is Sanders’ Equational Phonology. This one does differ from SPE principles on the matter of ordering (paragraph 2) and on the relevance of language use (paragraph 5), as well as on notational details coming under paragraphs (2) and (3). His position on the ideal speaker’s grammar (1) and on innateness (3) is not clear from this paper. Aside from the metatheoretical material in the earlier pages, Sanders’ most explicit claims have to do with unary features (i.e., features may be added or deleted, but absence of a feature—i.e., minus value—cannot be specified in the input), the use of general principles for settling conflicting rule application possibilities (similar to those mentioned by Anderson), and a claim much like the Houlihan-Iverson and Dinnsen ones (but broader) that all changes in final position (i.e., before juncture or silence) must be deletions, e.g., devoicing rather than voicing, segment loss rather than segment insertion. The same examples which violate their principles also violate Sanders’. (To make Dinnsen’s position resemble Sanders’ more closely, it is only necessary to insure that all atomic rules applying to final segments mention only + values in the input.)

In all three of these papers (Atomic, Functionally Constrained and Equational) it is either stated or implied that there cannot be rules of opposite sense (i.e. A → + B/C and A → — B/C__) either in different languages, different stages of one language, or simultaneously in the same language. Leben and Robinson (1977) cite an example from Schane attributed to Rumanian (where the environments are slightly different, but one is included within the other) and Sanders cites a Spanish example from Saporta, both of which are rejected. Pullum (1976) calls this (when the rules apply within the same language) the Duke of York gambit, and after a long discussion comes to the conclusion that there is no “basis for a general constraint that would prohibit” it: “The child is NOT equipped with a subconscious instruction” to avoid constructing grammars allowing Duke of York derivations.

From the counterexamples to all three of these papers, it is clear that the “universals” in question (regardless of how best to state them) are not true universals, comparable, e.g., to the claim that all natural languages have, on the phonetic surface, both consonants and vowels, or that no language lacks obstruents, or even the Chomsky-type universal that no language requires every word to be a phonetic palindrome. They are what we called “statistical universals” at Dobbs Ferry; propositions that are usually true of human languages, but whose failure to apply does not make a language incredible. They cannot, therefore, be part of the genetically transmitted LAD (unless we can assume mutations in some tribes), nor can they be logical necessities for any language implicit in its defining characteristics. They may, however, be accounted for in at least two different ways: either they make a language more efficient at whatever it is for (the factor considered irrelevant in SPE), or else they are due to primarily physiological causes but still do not make a language conspicuously less efficient.

If we turn now to Hooper’s Natural Generative Phonology, we find, again, differences that hardly warrant considering it a whole new phonology, but which are of some importance. It shares more than the name with Stampe and Donegan’s Natural Phonology: both make a sharp distinction between natural phonetic processes or P-rules and learned phonological or morphophonemic rules (MP-rules), the main difference being that Hooper goes further in claiming exceptionless unsuppress-ability for processes, while Stampe allows some of them to be overridden by rules acquired later. For Hooper a process is always surfacetrue, but she does allow a certain range of strength; it may be applied relatively more weakly or more strongly, resembling in this respect (at least) the low-level phonetic rules of SPE. Some P-rules are productive; others are unproductive, and this can be tested by their application to new loans or innovative creations. She does not, however, seem to make the distinction set up by Anderson between purely morphological rules (where, e.g., a change of one feature in itself signals a morpheme) and MP-rules. Much of her argumentation is based on speculations about how language-learners construct generalizations (including MP-rules), a kind of evidence nowhere evident in SPE, where the convention of paragraph (1) that language-learning is pretended to be instantaneous effectively cuts it out. Such arguments are also used by Stampe, Goldsmith, Leben and Anderson, and seem to me to be properly used in many cases. Hooper explicitly rejects the principle of paragraphs (3) and (4) that all forms of a derivational or inflexional paradigm must be phonologically derived from a single base form (root or stem), and allows for allomorphs in the lexicon. However, speculations about variants should pay more attention to the known facts, as when she couples the variant plurals hoofs and roofs, hooves and rooves (actually this conceals additional variants: the oo in all four forms may represent /uw/, but only in the first three /U/, at least in American dialects known to me). The status of the variants is quite different: both hoofs and hooves have been in spoken and written use for centuries, and for some of that time hooves has had a slight preference in most areas, but the form rooves has only become common in relatively recent years, and only in a few areas (and is almost unknown in printed usage), whereas roofs has been virtually standard everywhere for centuries.

Much of the presentation of Vennemann’s theory of rule inversion and semantic transparency should certainly have been linked up with Kuryfowicz (1949) on Analogical Change, but even Vennemann (1972a) does not so link it; many of the details are similar. Also, one form of the semantic transparency hypothesis may be amenable to testing by psycholinguistic experiments. Essentially the argument is that speakers will make use of every available phonological clue to decide an ambiguity; I have heard this doubted, however, and would very much appreciate an objective demonstration of some kind. Particularly the claims that “morphologically conditioned rules are never deletion rules” seems a bit strong. Even in English we have forms like has, had for [hævz], [hævd], which surely looks like a morphologically conditioned rule. Hooper’s discussion assumes an equal pairing of one against one, but if, as here, it is one or two forms with zero against many without, the reasoning of the language-acquiring speaker might conceivably be different. The celebrated case of the Russian genitive plural in zero must also be relevant here. Finally, I don’t think it can be considered to be self-evident to a language-learner (or a linguist) which semantic category in a paradigm is basic. Why did the Greeks opt for first-person singular present-tense as basic? Phonologically it won’t work, as they realized themselves.

Turning now to Donegan and Stampe, we find a great deal more concern with what SPE regards (in part rightly, I think) as low-level irrelevant surface phonetic variation. They apparently reject the whole enterprise of generative phonology (as one questioner complained, this is not a phonology at all) as set out above in paragraphs (2) and (4), and even conclude their paper by reinstating the autonomous or taxonomie phoneme, using very much the same argument as one I used myself in Householder (1965), the argument from native speakers’ awareness. However, as was shown in Saporta and Contreras (1960), for instance, native speakers are perfectly aware of some sub-phonemic matters. In the main the paper tries to make a contribution to the theory of language acquisition and of phonological performance, in both of which we have a natural emphasis on “processes” or P-rules. SPE is specifically contradicted as to point (5), much of point (1), almost all of (2), most of (3) (where a different set of devices are put in the brain of the language-learning child), and all of (4). But, in spite of this wholesale rejection, nothing much is put in the place of what is rejected. We are not told what exactly the function of the phonologist is, what the nature of the lexicon is or the shape of lexical entries, and very little about any early steps leading from the lexicon to performance, and what we get about the later steps is heavily anecdotal. There is a weak negative specification: “it is not intended to describe its subject matter exhaustively, i.e., to generate the set of phonologically possible languages,” but that doesn’t help much, since SPE nowhere, even in the famous chapter nine, claims to do that, though it may be somehow suggested that such a goal is desirable. Donegan and Stampe, unlike Hooper, deal more in tendencies than in exceptionless P-rules, and, since two tendencies may have opposite effects, the natural result is something like Anderson’s exegetic adequacy. The main novelty of Donegan and Stampe’s theory is the doctrine of language acquisition as suppression of natural substitutions (P-rules), selectively governed, I guess, by the child’s observation of the speech of older children and adults. They pay a great deal of attention (properly, I think) to phenomena of relaxed or allegro speech, but then grossly exaggerate the frequency of these phenomena (“the ordinary pronunciation of languages”). Anyone who takes the pains to listen carefully (play at half-speed or in reverse, if you like; make sound spectrograms of dubious items) to large quantities of spontaneous, natural speech will be surprised not by the “complex mappings of intended sounds and actual sounds” but rather by the relative rarity of these reductions and slurrings. Donegan and Stampe cite other bits of unsupported evidence, e.g., “speakers of many languages which lack final obstruents devoice these ... in foreign loanwords” (I don’t know of one such language and Donegan and Stampe do not name one); the American speaker who cannot pronounce final /l/ in English, but can in German because he is trying for a light /l/; the southern idiolects which interchange the vowels of bad and bite (reminding me of the Hoosier speaker who interchanges the vowels of barn and born or the German immigrant who interchanges /w/ and /v/); “most English speakers” who substitute [o?m] for [opm] (not me, at any rate; my glottal closure is simultaneous with the labial closure, not prior to it, and my velic release is well after the labial closure); the non-statistics on palatalization (or dipthongization) in bash as opposed to passed you. I am not alleging that any of these statements is false, merely that we have no way of checking and are offered no reason to believe that they have been checked. I also admire the Ob language in which all stressed vowels are merged, making painting, panting, punting and pointing identical; I am more familiar with Ug, in which the inserted syllable is unstressed. In Ug, however, betted and bedded might still be identical. The Middle Welsh change of /y/ to /ɨ/ looks very similar to the possible confusion of Azerbaijani /Ø/ with /ɯ/; it is quite easy to get the same acoustic effect with centralization as with rounding (and this is presumably what is meant by the expression “internal rounding” which occurs in the literature); see Householder (1972). Such a “two-step” change is like the familiar morphosemantic change of perfect to past which has occurred in so many languages of Europe, and is probably a legitimate exception to Austin’s law (1957).

Like Donegan-Stampe, Leben offers as “Upside-down” phonology something that can be called a phonology at all only by straining the meaning of the term. Everyone else at this conference seemed to agree with SPE that a phonology is a device for mapping lexical representations onto phonetic actualizations. Leben’s theory does not purport to do this. Instead, it attempts to satisfy the SPE concern over etymology in a different way, and answers this question: “If a native speaker ever happened to wonder if lexical item A was related to lexical item B, how could he go about testing?” Remember that for SPE phonology we have an idealized native speaker who intuitively knows all such relationships and has machinery to derive one of these items from the other if the relationship is real; he could never even pose Leben’s question. (Like Hooper and Donegan-Stampe, Leben is clearly rejecting this idealized speaker-hearer.) The method chosen is essentially a parsing algorithm: the speaker compares pairs of lexical entries with respect to one or more “morphological” rules (by this Leben means the phonological rules of SPE, at least in large part; and also the phonological rules of Anderson—perhaps with the addition of some of Anderson’s morphological rules-and surely Anderson’s “quasi-systematic relations”) and then reaches a decision as to relatedness. But two questions are left unanswered: (1) how can this algorithm be considered, in any sense, a part of the grammar of speaker’s competence? (2) who is going to ask such questions? Does a speaker who never asks them have a defective competence? Leben realizes that there will be some false positives, among which he mentions caustic-cost and comic-calm (apparently for him these contain the same surface vowel), and I have also thought of Cuba-cubic, fill-filth, terrible-tear, oar-oral and many more. To solve this problem, he stipulates that the parser must specify “word relationships that are linguistically possible.” But how on earth are we going to do that? This is the traditional bugbear of the historical etymologist: how much semantic similarity is sufficient for plausibility? And what native speaker-hearer, no matter how unidealized, is going to have access to appropriate criteria? There are really only two solutions to this problem: either (1) concede that any similarity which satisfies the native speaker is O.K.—which may well entail allowing mutually inconsistent pairings (and different for different speakers) simultaneously in the manner of Varro and other pre-Indo-Europeanist etymologists, e.g., relating filth independently to foul, fool, fill and perhaps feel, or (2) reject the criterion of semantic similarity entirely, which would unquestionably allow all the mutually inconsistent pairings just mentioned. Perhaps the most realistic is a third: accept the word of a big dictionary. This is certainly what real speaker-hearers often do. But however it is done, it clearly does not have to be done at all; a real speaker-hearer could have perfect command of his language without ever wondering about any such pairs, at least as far as Leben has shown.

So much for the papers in this volume. What might a stranger have expected of this conference? Several of the audience, in fact, told us about their disappointments. A glance at the SPE tenets (1-5) listed above shows very many opportunities for divergence and disagreement, phonological theories which would really be different theories. For instance, one could reject the notion of paragraph (1), striving instead for simple weak equivalence (the grammar produces the same output), using any conveniently intelligible notation and nomenclature without attempting to justify them by proofs from psycholinguistic experiments or otherwise. Automatically the rest of paragraph (1) would also be rejected. Even without that, one could follow Labov and others in rejecting (la), and trying to account for variation correctly. If you reject all or part of paragraph (2), you could then look upon it as part of the phonologists’ task (as it is the taxonomists’) to justify a particular surface segmentation (Is /ay/ in like one segment or two? Is long /ō/ in any language with a length contrast one segment or two? etc. Does the boundary between initial /p/ and a following stressed vowel in English begin at the moment of release, at the onset of voicing, or after the formant transition is complete?), or a particular surface “phonemicization.” Or else you could reject the notion of very long sequences of rules between lexicon and surface, requiring lexical entries to be less abstract (as, in part, Leben does, and possibly Donegan and Stampe or Hooper, though it is difficult to tell). Rejecting the innateness hypothesis might not, as Donegan and Stampe suggest, make any difference, but it might encourage people to look harder for functional or survival-value explanations of universals. Many of the details here have, as Anderson points out, already been rejected by most or many phonologists.

Paragraph (4) has already been rejected by several of the conferees; but if paragraph (1) is rejected, this sort of deep lexical structure could be reinstated in the manner of Pāṇini, allowing the convenience of many bits of abstractness which cannot be justified by a brain-correspondence theory. A linguist’s grammar (as for Pāṇini, Saussure and Bloomfield) could again become a worthy and respectable enterprise. SPE is, in many respects, such a linguist’s phonology, as is shown particularly by the idealized speaker-hearer and the instantaneous acquisition elements. Surely also we could modify the lexical entry (as I suggested in Householder 1971) by including orthography; this greatly simplifies many of Leben’s problems: you would not link caustic to cost because orthographic au and o rarely if ever correspond, but admission, adhesion, adhesive and the rest (not to mention telegraph and telegraphy) become less problematic. It is not that orthography corresponds so well to systematic phonemics (as Chomsky and Halle say), but that systematic phonemics has been created out of orthography; and many of the rules of SPE are modified rules of spelling-to-sound conversion. Several of the conferees have already rejected point (5), but more use can still be made of the functional aspect of native-speaker reactions. For instance, consider neutralization rules, which concerned several of the theoreticians. I cannot affirm what happens in all languages which have a rule of final devoicing when another speaker fails to apply it, but I can report the responses of speakers of two such languages (Turkish and German). (1) Failure to devoice goes quite unnoticed or else (if the speaker is a foreigner) receives praise. Many speakers flatly deny that they do devoice. (2) If an inherently voiceless final, however, is voiced, that is noticed; it either elicits criticism or blocks communication. (3) If an inherently neutral final (i.e., one which has non-distinctive voicing in some positions) is voiced, this also may go unnoticed. Neutralization does not seem to be wholly neutral.

One last possible reform that used to be much in the news involves restricting the “power of the theory,” i.e., making it impossible to do certain things. This was one of the motivations for chapter nine of SPE, which, as Anderson remarks, appears to be a really dead issue. Sanders offers some details to restrict the power slightly, but even in his scheme many impossible rules can be written. Dinnsen’s atomic rule proposal, likewise, involves only a slight restriction, and nobody else seems to have any proposals of this sort, though most of them would entail at least a reduction in the number of rules. Is this line of innovation really dead? Or would we not rather keep the powerful devices and concede that they are, after all, linguists’ devices that correspond to nothing particular in the speaker’s brain?