CHAPTER ELEVEN

On Arguing about Phonological Theories

Fred R. Eckman

1.0 During the general discussion session for this conference, there appeared to be some disagreement as to whether or not the proponents of the theories represented at the conference had provided any empirical basis for judging their theories to be superior to other current theories. The purpose of this paper is to sketch how it is that linguistic theories can be empirically differentiated, and to point out that, in fact, some of the papers did attempt to provide such a differentiation.

2.0 One of the principal goals of a theory of language is to provide a characterization of the notion “human language.” Most generative theories of language have gone about this by specifying a framework within which grammars of individual languages can be constructed. Such a framework has been referred to as a general linguistic theory, or linguistic metatheory. A grammar of some language constructed within any given metatheory constitutes a theory of that language. A metatheory, on the other hand, constitutes a theory about grammars, or equivalently, a theory about language. A metatheory attempts to characterize the notion “human language” by specifying what can and cannot be a grammar of a language.

In view of this it is clear that metatheories and grammars are theories about different types of objects, and therefore are supported or falsified by different types of evidence. A grammar can be falsified by showing that it is incapable of generating certain well-formed utterances in the language, or that it generates certain ill-formed utterances. A metatheory is falsified by showing that there is some language for which that metatheory fails to provide an adequate grammar, or that the metatheory provides grammars for objects which are not human languages. Consequently, metatheories are not shown to be false by falsifying a particular grammar, or a particular analysis, since the falsification of a particular grammar does not show that some other grammar, which is adequate, cannot be constructed within that same metatheory. What must be shown in order to falsify a metatheory is that the type of grammars that are constructed within that metatheory are inadequate, or alternatively, that although some adequate grammars can be constructed, the metatheory provides for the construction of grammars which are not grammars of human languages.

Conversely, a particular metatheory is not shown to be correct by showing that it provides adequate grammars for some language or some set of languages, since the possibility is left open that the metatheory provides grammars for objects which are not human languages. What must be shown, if the metatheory is to make any claim at all about the nature of human language, is not only that the metatheory correctly provides for adequate grammars of objects which are human languages, but also that it correctly excludes other objects as being impossible human languages. In other words, there must be some set of objects which the metatheory correctly excludes from the class of possible human languages by making it impossible to construct grammars for these objects within that metatheory.

Within this context, then, it is not sufficient for proponents of the individual metatheories represented at this conference to show that they can provide a simpler, more general or more elegant analysis of some set of facts. Rather, if these theories are to have any empirical import at all, they must show what it is that they correctly allow and what it is that they correctly disallow in terms of types of languages. In what follows, I will attempt to point out where some of the papers in this volume followed this reasoning explicitly and, therefore, provided an empirical basis for differentiating their particular theories.

It should be pointed out that the general conclusion that can be drawn from this is not altered significantly if it is the case, as was suggested in the general discussion session, that the theories represented at the conference are not really different metatheories, but only extensions or revisions of the Standard Theory presented in Chomsky and Halle (1968). Even if these theories do constitute merely revisions to the Standard Theory, they constitute revisions to a metatheory. That is, they represent changes in the form of grammars that can be constructed within this metatheory. In arguing for the revised form of a metatheory, one is arguing for the characterization of the notion human language that this revised metatheory provides. Thus, again, one must show what it is that this revised metatheory allows and disallows in the way of human languages.

3.0 The paper by Dinnsen (this volume) argues for the superiority of the theory of atomic phonology by showing that this theory explicitly excludes some types of rules from occurring in the phonologies of human languages. The central thesis of this theory is that the phonological component of human languages can contain only rules drawn from a set of atomic rules, or rules derived from these atomic rules by a principle of complementation. Thus, according to this theory, if the rule in (1) is an atomic rule, then the rule in (2) may not occur independently in the phonology of any language.

Consequently, the theory of atomic phonology excludes from the class of possible languages any language which has a rule which devoices only word-final fricatives. This theory makes the empirical claim that there will be no language which has a voice contrast in stops, a voice contrast in word-medial fricatives, but no voice contrast in word-final fricatives.

Similarly, the paper by Houlihan and Iverson (this volume) is an attempt to explicitly exclude language types from the set of human languages. Given their Markedness Constraint, it is claimed that it is impossible for the grammar of any language to contain a phonologically conditioned rule which is a neutralization rule and which converts only relatively unmarked segments into only relatively marked ones. Within this framework, it is claimed that there will be no language whose grammar contains an intervocalic voicing rule which is neutralizing. That is, there will be no language which exhibits a superficial voice contrast in some positions, say, word-initial and word-final positions, but exhibits only voiced obstruents in word-medial position.

Another example of a theory which shows how it can be empirically falsified by showing what it excludes as a possible language is the theory of equational grammar as put forth by Sanders (this volume). Within this framework, all grammatical rules must be expressed in the form of equivalence statements, and all principles governing the directionality, optionality and interaction of these rules must be universal. The universality of these principles, as Sanders explicitly states, excludes the possibility that the grammar of one language will have a phonological rule which is the converse of a phonological rule in the grammar of some other language. Thus, for example, if one language has a rule of terminal devoicing, there can be no language which has a rule of terminal voicing. This metatheory therefore predicts that there can be no language with only word-final voiced obstruents, some of which alternate with word-medial voiceless obstruents.

Having considered some examples of phonological theories which have shown how they can be empirically falsified by showing what it is that they explicitly exclude from the class of possible human languages, I would now like to briefly comment on some theories which did not take this approach.

The paper by Goldsmith (this volume) is an exposition of the theory of autosegmental phonology, which differs from standard generative phonology in that autosegmental phonology views phonological and phonetic representations as multilinear rather than as a single string of segments. Goldsmith argues for the superiority of autosegmental phonology by demonstrating the adequacy of this theory in characterizing the facts of a number of tone languages. However, Goldsmith does not try to show how it is that in characterizing these facts, the theory of autosegmental phonology also correctly excludes certain objects as being possible human languages. In fact, he takes the opposite approach, as shown by the following statement:

The reasoning behind this seems to lie in the assumption that increasing the conceptual framework of a metatheory enables one to gain certain insights which would otherwise be missed. What is not clear to me is how one gains more insight into the notion human language by increasing the class of possible grammars rather than by limiting this class. Hypotheses which pursue the latter approach are more easily falsifiable, and the metascientific value of the falsifiability of any hypothesis has been established, it seems to me, beyond question.

Another example of a theory which does not explicitly show what it excludes in terms of language types is the theory of natural generative phonology, as put forth by Hooper (this volume). In this paper, Hooper gives an extensive discussion of morpho-phonemic rules (MP-rules). Among the properties attributed to these rules are that they are morphologically conditioned, and the structural changes that they make are phonologically arbitrary. Thus, according to this theory, the phonologies of human languages can contain a set of surface-true, phonetically-motivated phonological rules (P-rules), and a set of MP-rules which are governed not by principles of phonology but by the Semantic Transparency Hypothesis. The consequence of this is that morphologically-conditioned phonological rules are excluded as possible rules in the phonologies of human languages.

So far Hooper seems to be proceeding in the direction of showing what it is that the theory of natural generative phonology excludes from the class of possible human languages, thereby making an empirical claim concerning what this metatheory characterizes as a possible language. Thus, she states near the beginning of her paper,

One such principle is the Semantic Transparency Hypothesis, about which she says,

Therefore, given that P-rules are governed by phonetic constraints, and that MP-rules are governed by the Semantic Transparency Hypothesis, it should be possible to determine what can and cannot be a rule in the phonological component of a language, and thereby make some claim about what can and cannot be a human language.

The problem with this is that it is not clear that anything is excluded from being a MP rule by the Semantic Transparency Hypothesis. Not only are the alternations governed by MP rules phonologically arbitrary, but the rules themselves are considered to be part of the arbitrary sound-meaning correspondence, similar to the principle which relates the semantic concept of a table to the phonetic string [mesa] in Spanish. Since we have no constraints (other then phonotactics) which govern what can and cannot be a possible sound-meaning correspondence, it would appear that nothing is excluded from the class of MP rules. This being the case, it is not clear what natural generative phonology excludes from the class of possible grammars, and therefore it is not clear what characterization natural generative phonology gives to the notion of human language.

4.0 The purpose of this discussion has not been to^. argue for or against any of the theories presented in this volume, since, as was pointed out in the discussion sessions, there are many unsolved problems. The fact that some of the papers pointed out how it is that the metatheory which they were espousing could be falsified obviously does not attest to that metatheory’s correctness; nor does the fact that other papers may not have shown what it is that the metatheory they were presenting explicitly excludes confirm the ultimate incorrectness of that metatheory. Rather, the purpose of this discussion has been to attempt to illustrate how it is that a proponent for any given metatheory should go about arguing for that metatheory.