IS EXPLANATORY ADEQUACY ADEQUATE?1
V. J. COOK
Version of paper in Linguistics, 133, 21-32, 1974
Though Chomsky's 'standard' theory of transformational grammar has been widely criticised, the criticisms have mostly been directed at the base subcomponent and the semantic component, allowing other aspects of the theory to be comparatively neglected. In particular, out of the three levels of adequacy outlined by Chomsky — observational, descriptive, and explanatory — only descriptive adequacy has received anything like the attention it merits (Peters, 1970; Seuren, 1970); explanatory adequacy, for instance, has featured chiefly as a digression in more general works (Wilks, 1972; Greene, 1972). The present article is an attempt to redress the balance by looking more closely at the crucial level of explanatory adequacy.
A convenient starting-point is the familiar diagram of language acquisition given in Fig. 1. On the left is the input, primary linguistic data consisting of samples of actual speech. On the right is the output, the generative grammar expressing the ideal speaker/hearer's tacit knowledge of his language, his linguistic competence. In between comes the Acquisition Model, the black box with unknown contents that relates input to output. The three levels of adequacy can be seen in terms of these three components (Chomsky and Halle, 1965). A linguistic description is observationally adequate if it correctly describes the primary linguistic data; it is descriptively adequate if it correctly describes linguistic competence. But the linguist may have the further aim of accounting for linguistic competence, of showing that the generative grammar is a product of certain properties of the human mind. "A linguistic theory meets the level of explanatory adequacy insofar as it succeeds in describing the internal structure of AM and thus shows how the descriptively adequate grammar arises from the primary linguistic data" (Chomsky and Halle, 1965: 100). For this to be possible, the theory must provide a reason for selecting the descriptively adequate grammar from the range of possible grammars; it must have an evaluation procedure — also confusingly known as a simplicity measure (Garcia, 1967). An evaluation procedure requires an hypothesis about the Acquisition Model and a method for assessing the extent to which a grammar reflects this hypothesis — an evaluation measure. An evaluation measure is not predetermined but "is itself an empirical hypothesis concerning universal properties of language; it is, in other words, a hypothesis, true or false, about the prerequisites for language-acquisition" (Chomsky, 1966: 22). An example of an evaluation procedure in action is the proposal that "The 'value' of a sequence of rules is the reciprocal of the number of symbols in the minimal schema that expands to this sequence" and that these symbols should consist of distinctive features (Chomsky and Halle, 1968; 334; Halle, 1961). It has been argued that on the one hand this provides a specific hypothesis about language acquisition and on the other suggests a method for evaluating grammars by counting the number of symbols, in this case distinctive features (Chomsky and Halle, 1965). An hypothesis such as this can be confirmed or disconfirmed. In essence this is done by comparing the input and output in Fig. 1: "To support or refute this hypothesis, we must consider evidence as to the factual relation between primary linguistic data and descriptively adequate grammars" (Chomsky, 1966: 22).
While these proposals fit neatly into the framework for explanatory adequacy given above, few other evaluation measures have been suggested. Chomsky himself has given some examples taken from syntax where he claims that explanatory adequacy is attained; at first it is hard to bring these within his own framework until one discovers that he has made "the assumption that the concepts of grammatical structure and 'significant generalisation' made explicit in this theory constitute the set of tools used by the learner in constructing an internal representation of his language (i.e. a generative grammar), on the basis of presented linguistic data" (Chomsky, 1964: 35). His examples are based on ascribing certain properties to the Acquisition Model and then measuring the extent to which alternative grammars reflect these properties. It is not at the moment relevant whether these hypothetical properties of the Acquisition Model are correct: visual perception could probably be analysed in terms of grammatical structure (Morton, 1971); the concept of the 'linguistically significant generalisation' may belong more to descriptive than to explanatory adequacy (Prideaux, 1971). It may well be that the specific hypotheses suggested by Chomsky and Halle could be disproved; this would not, however, demonstrate anything about explanatory adequacy itself.
Now there is no reason to suppose that human beings are born with different acquisition models; the language of the people the child encounters rather than that of his biological parents is the one he learns. Consequently the level of explanatory adequacy is concerned with linguistic universals rather than features peculiar to one language. A postulated property of the Acquisition Model with its associated evaluation measure is an hypothesis that applies to all languages: the relationship between primary linguistic data and generative grammar is the same regardless of the language involved.
While explanatory adequacy as summarised so far seems useful and workable, certain difficulties become apparent when it is examined more closely. In particular the concept of 'Acquisition Model' is deceptive. The term is in fact ambiguous, referring not only to the 'paper' model constructed by the linguist but also to the child's mind itself; this ambiguity is equivalent to Chomsky's use of 'grammar' to refer both to the linguist's representation and to the mental reality in the speaker's mind. It might seem that any property suggested for the paper model should also be found in some form in the child's mind: if the paper model has the property of preferring phonological rules with the smallest number of features, then so has the child's mind. Plausible confirmation for postulated properties of the Acquisition Model might come from evidence about how children learn language as well as from comparisons of input and output in the paper model. But the relationship is not so straightforward; for the Acquisition Model that Chomsky employs is instantaneous. "I have so far been discussing language acquisition on the obviously false assumption that it is an instantaneous process" (Chomsky, 1972; 89 fn.). This is amplified elsewhere: "A more realistic model of language would consider the order in which primary linguistic data are used by the child and the effects of preliminary 'hypotheses' developed in the earlier stages of learning on the interpretation of new, often more complex data" (Chomsky and Halle, 1968: 331). Any statement about acquisition within the framework outlined in Fig. 1 must be qualified: "If it were the case that language acquisition were instantaneous, then the underlying lexical forms with pre-Vowel-Shift representation would be psychologically real. This, we propose, is a true statement about language — ultimately about mental processes and the particular way in which they function" (Chomsky and Halle, 1968: 332). Inasmuch as the Acquisition Model relates to the child as well as to the paper construction it is concerned with an idealised child whose entire development takes place in an instant.
The justification for this 'obviously false assumption' is convenience: "To us it appears that this more realistic study is much too complex to be undertaken in any meaningful way today and that it will be far more fruitful to investigate in detail, as a first approximation, the idealised model outlined above" (Chomsky and Halle, 1968: 331). However, it can be argued that an instantaneous model is necessary because of the distinction between competence and performance. The factors that are excluded from competence but present in performance can be broadly classified as dialectal, situational, and "grammatically irrelevant conditions" (Chomsky, 1965: 3). The latter share some common features. Take memory for example. Memory limitations are excluded from competence; memory is necessary to performance because performance proceeds through time. The listener must store the earlier parts of the sentence while he is still hearing the later parts; since speech is not presented instantaneously memory processes have to be postulated. Similarly in speech production the fact that the whole sentence cannot be spoken at once necessitates a memory for planning ahead and for remembering what has already been said; the mistakes that are found in performance but excluded from competence are most likely caused by the involvement of speech production with time (Fry, 1969; Fromkin, 1971). Part of the idealisation of competence from speech processes is the detachment of competence from time. If competence is timeless and memory-less, then descriptions of competence cannot be evaluated by a measure that relies upon time and memory; otherwise the original point to the separation of competence and performance would be lost. In so far as competence is detached from time it forces the linguist into adopting a timeless evaluation measure and a timeless Acquisition Model, Though, looked at from one direction, the instantaneous model may be a convenient short cut, looked at from another, it is a consequence of the distinction between competence and performance. While Chomsky and Halle imply that the instantaneous model can expand to incorporate wider and wider evidence, the step from an instantaneous model to a non-instantaneous model may be more critical than they seem prepared to admit.
Whatever justification there may be for the instantaneous model, there still remains the problem of linking it to evidence about children learning language. If the model ignores stages in the child's acquisition, it can only take into account the global relationship between the total primary linguistic data and the final grammar of competence. For instance, it may be that children invariably start to speak by using holophrastic sentences derived from complex 'deep' structures (Bloom, 1970); this is irrelevant to the instantaneous model. It may be that all children then pass through a stage when, in terms of surface structure, their utterances consist of combinations of two word classes (Braine, 1963): this too cannot be taken into consideration. Even if both the content and the order of stages of acquisition were universal, this would still be beside the point. Chomsky and Halle's proposals about phonology would be in no way confirmed if it were shown that each stage in the child's acquisition had to be described in terms of features and that each successive stage economised on the number of features in some way. Restricting the model to dealing with global input/output relations means that one class of universals — universal stages of development — is specifically excluded. Compared to a theory such as Piaget's that makes strong claims about universal stages of development this instantaneous model is weak. Take for instance the ordering of rules. It might seem obvious that there should be some correspondence between the ordering of rules in competence and their order of acquisition, that a possible evaluation measure would select grammars in which the order of rules were closest to the order of acquisition, other things being equal. Children seem to learn English Wh-questions by starting with 'occasional' questions and then learning first a 'preposing' transformation and second a 'transposing' transformation (Brown, 1968). But without relaxing the requirement of instantaneity, this is not relevant: the Acquisition Model can contain no information about order and so cannot measure competence in terms of order. The instantaneous model has in effect cut itself off from evidence about the child.
Let us take an example from syntax to see what these criticisms amount to. Ross has proposed that all English sentences are derived from underlying structures in which there is a 'performative main clause' (Ross, 1970); in most cases this performative clause is deleted from the surface structure. Thus underlying Prices slumped is an underlying structure paraphrasable as I declare to you prices slumped. Ross supports this proposal with fourteen syntactic arguments of debatable validity (Matthews, 1972). Let us suppose that an alternative description of English exists that is identical except that it does not derive every sentence from an underlying structure containing a performative clause. To reach explanatory adequacy the theory must provide a method for selecting one of these grammars; it needs a suitable evaluation measure. Construction of this measure depends upon making an hypothesis about the Acquisition Model. The first step is to claim that the performative clause is a possible linguistic universal, as Ross suggests. We next postulate that the Acquisition Model contains a Performative Clause Property (PCP) which explains this. Now, how do we confirm that PCP exists? We might start by looking at the primary linguistic data; we find no ubiquitous performative clause there. We turn to developmental evidence about children; whatever we find there is, however, ruled out. Our next appeal is to the grammar of competence; since we have two rival grammars of competence that differ precisely in this feature, there is still no confirmation for PCP. Now we try comparing the primary linguistic data with the grammars of competence; again this is indecisive since in the one case PCP is confirmed, in the other disconfirmed. Within the framework of Fig. 1 we can find no evidence for or against PCP. The only proof of its existence is its presence in one of the rival grammars. But this was the reason for postulating PCP in the Acquisition Model in the first place. Hence it is not confirmed. Any evaluation measure we set up will indeed select one of the grammars: an evaluation measure derived from a performative analysis will prefer performative analyses. But it will provide no reason for this preference. The argument is circular because PCP can be checked neither against developmental data nor against the comparison of input and output. Setting up a property of the Acquisition Model on evidence from competence and then using that property to evaluate the evidence for its own existence seems rather like having your cake and eating it. This circularity cannot be avoided with any other property of the Acquisition Model that could be postulated: the existence of properties that prefer linguistically significant generalisations or distinctive features is vouched for only by their presence in the grammar of competence, in other words only by arguments of descriptive adequacy. Explanatory adequacy has not provided a principled method of evaluation additional to those employed in achieving the other levels of adequacy.
Can these objections be met by reformulating the model? The simplest change would be to adopt a non-instantaneous model, with the consequent redrawing of the competence/performance distinction: while this has been frequently attempted by researchers in allied fields (Labov, 1969; Moravscik, 1971; Watt, 1970), linguists seem to have left it well alone. Let us then see whether a non-instantaneous model such as that in Fig. 2 can overcome the objections. The first line in the diagram represents the child's initial state: he hears a certain amount of primary linguistic data; he applies to it his initial methods of processing speech, Performance ModelI; the output from this is processed by his initial methods of language learning, Acquisition ModelI; this in turn produces the first grammar of competence, Generative GrammarI. Then the child hears more speech, processes it with a slightly different Performance Model (which presumably draws upon Generative GrammarI); the output is processed by a slightly different Acquisition Model and leads to a more developed Generative Grammar. And so on for all the stages necessary to reach the terminal adult competence, Generative GrammarT. In the revised model both Performance and Acquisition Models intervene between the primary linguistic data and the generative grammar; both of these include memory processes as they are now involved in time, the Performance Model at least those short-term processes necessary to perception, the Acquisition Model at least those necessary for long-term memory (Craik and Lockhart, 1972); both of them are also continuously evolving rather than necessarily passing through a discrete number of stages. The revised model does not, however, preclude the possibility that the Performance and Acquisition Models should be collapsed into one (Bever, 1970).
The relationship between horizontal lines is then progress through time. Performance ModelI may process language in different ways in later Performance Models; this allows research on the qualitative development of memory to be incorporated in the model (Mehler & Conrad, 1971). Similarly the Acquisition Model may change through time: 'contextual generalisation' may be a property of Acquisition ModelI but not of subsequent models (Schlesinger, 1967). Changes in primary linguistic data can also be included; early stages may contain phonological and syntactic features not present later (Blount, 1971; Snow, 1972). Finally the development of the grammars of competence be accommodated (Klima and Bellugi, 1966). The question of universals has now been broadened to include possible universals of performance, of acquisition, and of sequence. The reduction transformation in early stages of syntactic development may be a performance universal (Bloom, 1970); the phonological development of babbling into a consonant/vowel opposition and the subsequent splitting of consonants nasal/oral and labial/dental, and of vowels into wide/narrow may be a universal sequence (Jakobson, 1968). Thus the non-instantaneous model is capable of unifying the types of grammar studied by linguists with the models of speech perception and language development studies of psycholinguists.
How would explanatory adequacy work in the revised model? Let us take up performative clauses once again, noting that Ross himself presents an alternative 'pragmatic' analysis that implies a breach of competence and performance. We have two rival grammars for position of GrammarT, one of them containing the performative analysis. As before, we propose that the Performative Clause Property (PCP) is part of the Acquisition Model and try to confirm this. First we check Acquisition ModelI to see whether it is present; then we check back to Performance ModelT and Primary Linguistic DataT. If it has not been confirmed so far we move back through earlier stages of generative grammars, acquisition models, performance models, and primary linguistic data. And so on till we either find the source of PCP or we reach Primary Linguistic DataI. At all points we can check the presence of PCP against evidence from the child's development: we have a direct link between our postulated universal and the child. Since everything does not have to be explained in terms of one global process it is also possible some aspects of acquisition that would otherwise be inexplicable can be explained in terms of an evolving multiple process model. Once PCP is confirmed we can set up an evaluation measure that prefers grammars containing performative analyses; the evaluation measure is no longer circular since it measures the alternative versions of GrammarT by other criteria than those derived from the grammars themselves.
This revised model has countered one of the objections to explanatory adequacy by linking the generative grammar to evidence from language acquisition. To see whether it has successfully countered the other objection, let us look at the case in which the existence of PCP has been checked all the way back to Primary Linguistic DataI but has nowhere been confirmed. One might conclude that it was disconfirmed and that an evaluation measure should be chosen that prefers grammars without performative analyses. But this conclusion is wrong. For the theory permits the existence of innate ideas, of properties of language intrinsic to the acquisition model. Rather than being disconfirmed, PCP is now a possible innate universal. Again the only evidence for its existence is one of the alternative grammars of competence; again a postulated property is being used to evaluate its own hypothetical existence. If the sole justification for an innate linguistic universal is its presence in competence then it cannot be used as a touchstone to measure grammars of competence in which its own existence is disputed. PCP might be used to evaluate competence if it had an independent justification; PCP might be deduced from competence if the evidence were sufficiently compelling; PCP cannot evaluate grammars in which the performative analysis is itself at stake. Properties of the mind can be deduced from pre-established grammars and from the relationship between primary linguistic data and pre-established grammars; without confirmation from other sources, postulated properties cannot be used as reasons for preferring one grammar to another.
Explanatory adequacy does not then seem adequate. The evaluation measure only measures global input/output relations separated from developmental data; when this is remedied, the evaluation measure is still circular to the extent that the theory permits innate universals. If properties of the mind are allowed to be innate, their existence cannot be confirmed from grammars of competence when their postulated existence was itself the basis for selecting the grammars in the first place. The linguist is then faced with the extremely difficult problem of confirming innate linguistic universals. Until some solution is found it is fruitless to talk about innate universals within a theory that still uses, even if tacitly, the approach to explanatory adequacy suggested by Chomsky.
Bever, Thomas G.
1970 "The Cognitive Basis for Linguistic Structures", Cognition and the Development of Language, ed. by John R. Hayes (New York, John Wiley), 279-362.
1970 Language Development: Form and Function in Emerging Grammars (Cambridge, Mass., M.I.T.).
Blount, Ben G.
1971 "Socialization and Pre-Linguistic Development in the Luo of Kenya", Southwestern Journal of Anthropology 21, 41-50.
Braine, Martin D. S.
1963 "The Ontogeny of English Phrase Structure: The First Phase", Language 39, 1-13.
1968 "The Development of Wh Questions in Child Speech", Journal of Verbal Learning and Verbal Behavior 7, 279-90.
1964 Current Issues in Linguistic Theory (= Janua Linguarum, series minor 38 (The Hague, Mouton).
1964 Aspects of the Theory of Syntax (Cambridge, Mass., M.I.T.).
1964 Topics in the Theory of Generative Grammar (= Janua Linguarum, series minor 56) (The Hague, Mouton).
1972 Language and Mind, enlarged ed. (New York, Harcourt, Brace, Jovanovich)
Chomsky, Noam, and Morris Halle
1965 "Some Controversial Questions in Phonological Theory", Journal of Linguistics 1, 97-138.
1968 The Sound Pattern of English (New York, Harper and Row).
1971 "The Chronology of the Development of Covert Speech in Children” Developmental Psychology 5, 398-405.
Craik, Fergus I. M., and Robert S. Lockhart
1972 "Levels of Processing: A Framework for Memory Research", Journal of Verbal Learning and Verbal Behavior 11, 671-84.
Fromkin, V. A.
1971 "The Non-Anomalous Nature of Anomalous Utterances", Language, 27-52.
Fry, D. B.
1969 "The Linguistic Evidence of Speech Errors", Brno Studies in English 8, 69
Garcia, Erica C.
1967 "Auxiliaries and the Criterion of Simplicity", Language 43, 853-70.
1972 Psycholinguistics: Chomsky and Psychology (Harmondsworth, Penguin
1961 "On the Role of Simplicity in Linguistic Descriptions", Proceeding Symposia in Applied Mathematics XII, 89-94.
1968 Child Language, Aphasia, and Phonological Universals (= Janua Linguarum series minor 72) (The Hague, Mouton).
Klima, E. S., and U. Bellugi
1966 "Syntactic Regularities in the Speech of Children", Psycholinguistic Papers: The Proceedings of the 1966 Edinburgh Conference, ed. by J. Lyon & R. J. Wales (Edinburgh University Press), 183-219.
1969 "Contraction, Deletion, and Inherent Variability of the English Copula", Language 45, 115-62.
Matthews, P. H.
1972 Review of Readings in English Transformational Grammar, ed. by Roderick A. Jacobs and Peter S. Rosenbaum, Journal of Linguistics 8, 125-36.
1971 "Studies in Language and Thought Development", Language Acquisition: Models and Methods, ed. by Renira Huxley and Elizabeth Ingram (London, Academic Press), 201-29.
1971 "Competence, Creativity and Innateness", Philosophical Forum I, 407-37.
1971 "What Could Possibly be Innate?", Biological and Social Factors in Psycholinguistics, ed. by John Morton (London, Logos), 82-97.
1970 "Why there are many 'Universal Bases'", Papers in Linguistics 2, 27-43.
Prideaux, Gary D.
1971 "On the Notion 'Linguistically Significant Generalisation'", Lingua 26, 337-47.
Ross, John R.
1970 "On Declarative Sentences", Readings in English Transformational Grammar, ed. by Roderick A. Jacobs and Peter S. Rosenbaum (Waltham, Mass., Ginn), 222-12.
Schlesinger, I. M.
1967 "A Note on the Relationship between Psychological and Linguistic Theories", Foundations of Language 3, 397-402.
Seuren, Pieter A. M.
1970 "A Note on Descriptive Adequacy", Journal of Linguistics 6, 263-66.
Snow, Catherine E.
1972 "Mothers' Speech to Children Learning Language", Child Development 43, 549-65.
Watt, William C.
1970 "On Two Hypotheses Concerning Psycholinguistics", Cognition and the Development of Language, ed. by John R. Hayes (New York, Wiley), 131-220.
1972 Grammar, Meaning, and the Machine Analysis of Language (London, Routledge and Kegan Paul).
1. I should like to thank the following persons for their helpful comments on earlier drafts of this paper: P. Meara, G. Pullum, E. Valentine, D. Whiteley.