CHAPTER FOURTEEN

Comments

James D. McCawley

Editor’s note: The title of this volume was to have been Current Phonological Theories. In response to McCawley’s remarks at the conference and reiterated herein, the title was revised.

The plural in the title Current Phonological Theories is misleading, since it wrongly suggests that each of the nine major presentations at this conference dealt with a separate phonological theory. In general, the various presentations dealt with specific issues in phonology, with little overlap among the issues that different presentations took up. The disagreements among the participants appear to be not so much disagreements over specific issues as disagreements as to what issues are most worth discussing. Thus, it is hard to see how there could be any conflict between Goldsmith’s position on the nature of phonological structures, Leben’s position on the ‘directionality’ of derivations, Dinnsen’s position on how the class of possible phonological rules is constrained, and Houlihan and Iverson’s position on the relationship between the function of rules and the allophonic/morphophonemic distinction. Except for Donegan and Stampe’s paper, which expounds a really comprehensive theory of phonology,¹ and to a lesser extent also Hooper’s paper, the various papers presented not phonological theories but treatments of particular aspects of phonology that could in principle serve as parts of a broad range of possible phonological theories.

I will devote the remainder of my remarks here to taking up certain points raised by the three papers that I was specifically asked to discuss. My neglect of the other papers should not be taken as implying that I did not find them stimulating—only that the three covered here are the ones that I have read the most carefully.

I wish Leben had dealt with an important issue that has considerable significance for his approach, namely that of how one can tell what morphemic relationships exist and which of those relationships are mediated by rules. I have commented elsewhere (McCawley 1976, 1977) on the failure of generative phonology to provide any basis for determining when two words share a morpheme (e.g., Does any variety of generative phonology provide grounds for saying that sign and signature begin with the same morpheme but moth and mother do not?) or to consider the possibility that different morphemic identities may have different psychological status. Leben treats his ‘upside-down’ derivations as comprising the whole of phonology, rather than considering the possibility that they might correspond only to a specific kind of morphemic relatedness. Since Leben’s derivations allow rules to ‘apply’ or not, depending only on whether their application allows for greater similarity in the underlying forms of the putatively related morphs, the most obvious function that they fulfill is simply that of an aid to perceiving the morphemic relationship in the first place: the rules simply tell one that certain differences between morphs can be ignored in deciding whether they can be identified. To say that there are rules having this function is not to say anything about whether there are other systems of rules with other functions. Leben’s derivations provide an eminently reasonable proposal for how language-particular phonological rules could play a role in the perception of morphemic identity. However, I find it highly implausible to suppose that all phonological rules have the sort of optionality that his rules do, i.e., that they can apply or not, whichever makes the deeper representations of putatively identical morphs more alike. Thus I would find his treatment of Christian, etc. far more plausible if he were taking it as typical only of non-productive morphology and were taking productive phonological processes as subject to a different set of principles of rule application. I of course do not rule out the possibility of there being an overlap between the rules of productive phonology and the rules that figure in the identification of non-productively related morphs, and there would be considerable interest in the question of what learning in the one domain might be carried over into the other.

I should add that it is doubtful that all psychologically real morpheme identifications can be covered by rules (taking a ‘rule’ to be something that in principle could figure in derivations other than the one given). Whatever general facility human beings have for perceiving similarity surely plays a role in the perception of morpheme identity,² and I conjecture that there is nothing specifically phonological in one’s perception of such morphemic relationships as that between [kīhṓwtēy] and [kwìksā́tik]. But what morphemic relations are learned on the basis of prior knowledge of rules, and how much uniformity is there in a linguistic community with regard to either what non-productive rules the various speakers know or what morphemic relationships they mediate? The notion that an orthodox generative phonologist is likely to invoke here, namely that of ‘evaluation measure’, is a red herring. An evaluation measure is supposed to provide a rate of exchange between rules and examples: each rule has a ‘cost’ that is offset whenever the number of examples to which it applies exceeds a certain threshold. But it gives no clue as to what examples should count as instances of a particular rule, e.g., it gives no clue as to whether cholera is a derivative of coal, with a derivation involving trisyllabic laxing. Moreover, as Stampe has pointed out, it is extremely implausible to suppose that on learning his n-th word to which a given rule would apply and thereby becoming able to afford the rule, a child rewrites his dictionary entries for the n-1 examples he had known before. The perception of morphemic relationships thus will often precede the learning of rules that could be taken as mediating the relationships. In McCawley (1976, 1977), I argue for a non-deterministic picture of language acquisition: there are numerous factors, some linguistic and some non-linguistic, affecting the likelihood of one’s perceiving a morphemic relation or learning a rule, and the only relationship between the number of instances of a rule and whether one learns it is that the more chances you have to learn a rule, the greater the likelihood that you will learn it. I accordingly would expect that there would be a great amount of interspeaker variation with regard to the status of the rules and the examples that Leben cites: variation as to whether a speaker perceives two given words as related at all, variation as to whether a morphemic relation that he perceives is connected with any phonological generalization that he knows, and variation as to what exactly the phonological generalization is.

Anderson’s discussion of curly brackets conflates two issues that should be kept apart: the issue of whether there are ‘rules’ that can be shown to function as units but which subsume separate cases that must be listed individually as in the curly bracket notation, and the issue of whether, if such rules exist, the curly brackets function as an ‘abbreviatory device’. That these are separate issues can be seen from my argument (McCawley 1974) that there are rules appropriately formulated in terms of ‘optional’ segments but that the notation (parentheses) used in formulating such rules is not an ‘abbreviatory device,’ since the rules that the notation supposedly allows one to put together need not be possible phonological rules, e.g., a voicing assimilation rule requiring the affected and conditioning segments to be separated by a glide is not a possible rule, though a voicing assimilation rule allowing an optional glide between the two segments is possible. I in fact am now inclined to agree with Anderson that Finnish consonant gradation is appropriately described in terms of a rule that gives the environment in which gradation takes place and lists (in curly brackets) all the different changes that gradation comprises.³ However, the plausibility that I see in the gradation rule that Anderson describes comes from a conjecture as to how consonant gradation might be acquired, and under this hypothetical picture of language acquisition, the curly brackets would not be an abbreviatory device, since the child would learn the curly brackets without ever learning the rules that the curly brackets supposedly abbreviated. Specifically, I conjecture that Finnish children first learn that strange things happen at the beginning of a short closed non-initial syllable and that only gradually do they learn the details of those strange things. The child would first learn a rule that could be formulated with an empty pair of curly brackets, and only later would he learn what went inside those curly brackets.

Under this conjecture, the Finnish child’s first gradation rule would be something that generative phonology has never countenanced: a rule that does not say what happens. Nonetheless, such a rule could play an important part in the child’s language use: it would identify segments in certain environments as potentially variants of other segments and would thus make many morpheme identifications more readily accessible to him than they otherwise would be. I note in this regard that under Leben’s approach there is no reason why rules in adult language might not also be incomplete. For example, I do not find it at all implausible to suppose that many speakers of English recognize morphemic relationships in terms of a rule V→ [?] (i.e., something happens to a vowel), a synchronic analogue of Voltaire’s famous remark about etymology.

The notion of ‘exegetic adequacy’ that Anderson introduces is really not a property of theories but rather of ‘research programs’ in the sense of Lakatos (1970). A research program is a set of policies accepted within a scientific community (or subcommunity) on what sorts of questions members of that community should ask and what sorts of answers they should seek. It is possible for different theories to be combined with nearly identical research programs, since one can hold that it is worthwhile to seek first a certain type of answer without necessarily subscribing to a theory that says that only that type of answer can be right; for example, you don’t have to believe that sound change is conditioned only by phonological factors in order to believe that an analysis in terms of phonological factors should be sought first.⁴ Many ostensible defenses of theories are really defenses of research programs. For example, when Bloomfield pointed with pride to the Swampy Cree data that confirmed his earlier reconstruction of Central Algonkian and when Chomsky (1972:198) spoke of the policy that “the formal devices of language should be studied independently of their use” as having been “fruitful,” their remarks really said nothing about theories. In both cases their remarks amounted to the claim that many valuable results have been achieved when a certain type of answer has been sought. However, the proposition that many valuable results have been achieved in no way conflicts with the proposition that many valuable results have been systematically avoided and that many highly dubious results have been achieved. Testimonials to a research program are comments on the author’s tastes rather than on any theory that may be attached to the research program: they say that the results achieved through the research program have been enough to keep him content, but they provide no basis for evaluating theoretical claims (such as the claim that sound change is conditioned only by phonological factors). It is misleading for Chomsky to describe the policy of studying formal devices independently of use as a ‘hypothesis’: a hypothesis says how the world is, not how researchers should investigate the world.

Goldsmith’s autosegmental phonology allows one to preserve what was worthwhile in Lightner’s (1965) now discredited analysis of vowel harmony, while avoiding the faults that led to its being discredited. Lightner treated Classical Mongolian vowels as underlyingly unspecified for gravity and had a morpheme feature [+/ — GRAVE] as part of the dictionary entry of each root. The [+ / — GRAVE] specification of a root conditioned the adding of its lower-case counterpart (i.e., [+ / — grave]) to all vowels that either belonged to the root or belonged to affixes attached to the root. Lightner’s proposal gives no clue as to why, in Finnish, Hungarian, and Turkish, when vowel harmony is violated within a root, suffixes normally harmonize with the last non-neutral vowel of the root: if harmony is conditioned by a morpheme feature, there is no reason why the last vowel of a root should have any more privileged position as a conditioner of vowel harmony than does the first vowel. The autosegmental treatment allows roots that conform to vowel harmony to have a single specification for gravity or whatever the harmonic feature is, and does not require the proliferation of features that Lightner’s treatment does (GRAVE ≠ grave), and the same principles that insure the spread of the single gravity feature onto both root and suffix vowels when the root conforms to vowel harmony will also insure that the last gravity specification spreads onto the suffixes when the root does not conform to harmony: all vowels have association lines, and association lines may not cross.⁵

There remain some problems with the autosegmental treatment of vowel harmony, however, particularly with regard to ‘neutral vowels’. It is normally the last non-neutral vowel, not the last vowel, that conditions harmony, e.g., Finnish dityrambi-na/*-nä, where the suffix harmonizes with the /a/, skipping over the ‘neutral’ /i/ (/e/ is also ‘neutral’ in Finnish vowel harmony). Nonetheless, neutral vowels are not irrelevant to vowel harmony, since when a Finnish root contains only neutral vowels, the suffixes take front-vowel harmony, e.g., kiel-tä, partitive of kieli ‘language,’ evidently reflecting the frontness of neutral vowels. If neutral vowels are taken to have a frontness specification on the front-back tier of an autosegmental structure, then such forms as pankki-na ‘as a bank’ will involve crossing association lines: the line connecting the backness specification of the root to the ending will cross the line connecting the /i/ to its frontness specification. Lloyd Anderson (1975) discusses a number of instances in which vowel harmony indeed crosses over non-neutral vowels, e.g., ambassadööri-na ‘as an ambassador’, in which the suffix harmonizes not with the neutral /i/ nor the front /öö/ but with one of the preceding /a/’s. Anderson’s solution to the problem of identifying the vowel with which a suffix harmonizes was to isolate a number of factors that affect vowel harmony and give an algorithm for determining which factor would prevail when they conflicted. Proximity was one of the factors (i.e., other things being equal, a vowel will harmonize with the vowel that is closest to it). Another was vowel quality: Anderson set up a hierarchy of strength with which different vowels condition vowel harmony, with /a/ at the top of the hierarchy, /ö/ in the middle, and /i,e/ at the bottom. In pankki-na, the strength of the /a/ as a conditioning factor overrides the proximity of the /i/, and in ambassadööri-na it overrides the proximity of the /ö/ and /i/. In words like kiel-tä there is nothing to override the proximity of the /e/, and the suffix takes front vowel harmony. Anderson’s approach to vowel harmony could be recast in revisionist autosegmental terms by making the principle against crossing association lines relative rather than absolute: various segments and/or association lines would provide obstacles of different strength to crossing association lines rather than presenting impenetrable barriers.

I dispute one universal claim that Goldsmith makes, namely that “there are no association lines in the stored or underlying representation.” In fact, as far as I can see, Goldsmith’s (1975a) own treatment of Japanese accent involves “association lines in the underlying representation,” in that the accent mark that Goldsmith allows in his underlying representations for pitch accent systems amounts to an association line: it indicates with which syllable a particular tone of the melody will be associated.

Finally, I would like to throw out a rather programmatic suggestion for autosegmental phonology, namely that the different tiers as well as the associations among them may be represented not only in discrete but also in continuous terms. In Goldsmith’s present framework, tones must be associated with vowels (and not with interstices between vowels), and a sequence of tones must be realized as a contour tone. It is thus impossible to represent a downstep as an unrealized low tone between two high tones, and Goldsmith is forced to adopt a “Low-high simplification rule,” which creates a “drop tone” out of an LH sequence associated with a single’ vowel. Suppose, however, that an HLH sequence were associated with two syllables in such a way that the first H was associated with the entire first syllable, the second H with the entire second syllable, and the L with the boundary between the two syllables, and suppose that the drop from an H to an L exceeds the rise from an L to an H. This association would cause the L to be unrealized as such but would leave it as part of the phonological structure and allow its presence to be manifested in the superposition of a small rise (L to H) on a large drop (H to L), which yields a small drop, i.e., downstep, at the syllable boundary in question. Such a generalization of the notion of “association” allows one to treat what Goldsmith calls “drop tones” as simply H’s (which is in fact what they behave like in Igbo phonology) without having to resort to representations in terms of absolute pitch values as in the derivations suggested in such works as Schachter and Fromkin (1968):

H L H

→  5 2 4 (numerical indication of pitch level)

→  5 4 4 (by assimilation) =H ’H H

The absolute pitch levels that figure in the inputs to the assimilation and deletion rules that give rise to downstep in Schachter and Fromkin’s analyses are highly questionable, in that it is doubtful that any rule would be sensitive to specific pitch levels, e.g., there can’t be a rule that raises all 2’s (regardless of whether they are realizations of H’s or of L’s), though there can be a rule that raises an L to the pitch of a following H. The extended autosegmental treatment that I have just suggested allows one to have essentially the same assimilation and deletion rules that Schachter and Fromkin do (the assimilation rule extends leftward the domain associated with an H) and yet do the entire phonology with only L and H tones.