JOURNAL OF THE EXPERIMENTAL ANALYSIS OF BEHAVIOR

A FUNCTIONAL ANALYSIS OF LANGUAGE'

DAVID PREMACK

UNIVERSITY OF CALIFORNIA, SANTA BARBARA

Language has been given a largely structuraldefinition by linguistics, but in order to havea psyclhological theory of language, the struc-tural emphasis must be replaced by a func-tional one. What must an organism do in or-der to give evidence that it has language? Morespecifically, when is a response a word? A se-quence of responses a sentence? What makesone response sequence an assertion or predica-tion, another an imperative, still another aquestion? In this paper I try to give thesequestions the most general answers possible,general in the sense of relieving them of theirexclusively human form.The functions an organism carries out when

engaged in language need to be separatedfrom the form these functions take in man.Not only human phonology but quite possiblyhuman syntax may be unique to man; bothmay encompass mechanisms not found in anyother species (Chomsky, 1965; Lenneberg,1968). But if this is so, it does not commit themechanisms of logic and semantics to thesame status. The latter may be more widelydistributed and it may be them, not the hu-man form of syntax and phonology, uponwhich the basic functions of language depend.

Strict Training Procedure: A Recipe forTeaching Language FunctionsThis paper is organized around two inter-

locking lists. The first is the list of functions,the things an organism must do in order togive evidence of language. The ideal list ofthis kind will be exhaustive, although pres-

1Based upon an invited APA address (Div. I), 1969,Washington, D.C. Mary Morgan, J. Olson, Randy Funk,and Deborah Petersen are the exceptionally patientand ingenious research assistants who made the studypossible. The research was supported by NIH grantMH-15616. Reprints may be obtained from the author,Dept. of Psychology, University of Califomia at SantaBarbara, Santa Barbara, California.

ently it is acceptable simply if it avoids glar-ing lacunae and contains only items that arepatently important. The second is a parallellist of strict training procedures. For each func-tion on the first list, the second gives at leastone and preferably several alternative ways ofproducing the function. A strict training pro-cedure is'essentially a recipe. Given a decisionas to what constitutes, for example, compe-tence in the interrogative-the ability to askand answer questions-the second list offers aset of instructions showing how to train an or-ganism so as to instill the competence in ques-tion. Clearly the difficulties of the first listgreatly exceed those of the second. It is farmore elusive a task to explicate what interro-gation or predication consist of than, given adecision on the former, to produce a trainingprocedure that will inculcate the competencein question. Indeed, a strict training procedureis no more than an ordered series of steps,each one to be accomplished before the nextone is begun, and each one so small as to beatomic, i.e., manageable by a docile organism.Although certain species, man notably andeven other higher primates, may be capableof taking many steps at a time-so that forthem an atomic decomposition of the task wasnot a necessity to begin with-the strict train-ing procedure is happiest when it reaches thelowliest organism; it knows no other way ofdoing this than by breaking the task into thesmallest steps possible. It should be recognizedfrom the beginning, however, that there is nomechanical procedure for generating recipes.The notion of an atomic step is a primitive.Furthermore, a strict training procedure is notan explanation of how, as a result of carryingout the prescribed steps, the organism ac-complished the function in question. A recipeis a method, not a theory, though by merit ofits success, often a method in search of atheory.

107

1970, 14, 107-125 NUMBER I (JULY )

DAVID PREMACK

Chimpanzee as a Drawing BoardThe chimpanzee will be our drawing board.

I will take only four items from the first list,which elsewhere I am attempting to deal withcomprehensively, and show how applicationof the corresponding training procedures tothe chimpanzee results in the functions inquestion. The functions considered are: word,sentence, question, and metalinguistics. Eachone opens out into further topics, some ofthem classical such as displacement-talkingabout things that are not present-and predi-cation-asserting a state of affairs (as opposedto requesting it). Displacement and predica-tion have both been cited as uniquely human,as hallmarks of man and language. Certainlythey are impossible without language but, as

we will see, they are not unique to man. It isunfortunate in a sense to use the chimp as

the drawing board, for it is too close to man.

It will be more illuminating if later the same

functions can be instilled in nonprimates. Butteaching an organism language amounts inpart to mapping the built-in knowledge of itsspecies; this will be brought out here in a

number of examples. And this knowledge isdifficult, even perhaps impossible, to disinterin species far removed from man. The greateraccessibility of this knowledge in the case ofthe chimp is the main reason we start with it,more than the fact that it is bright andplayful.

Physical Basis of Language: Plastic Words

The physical basis of the language used withthe chimp is plastic varying in shape, size,texture, and color. Each word is a metal-backed piece of plastic that adheres to a mag-

netized slate (see Fig. 1). The sentences are

written on the vertical, an ancient form ofwriting once used by certain human groups

(Hewes, 1949), but adopted here simply be-cause in the beginning it appeared to be thechimp's preferred style. The two sentencesshown in Fig. 1 can be paraphrased in Englishas follows: "Sarah take honey-bread," and "noSarah take jam-cracker," respectively. Sincethe language is written rather than spoken or

gestured, words are permanent not evanescent,and sentences are displaced in space not time.This has overwhelming advantages for short-term memory. Once written on the board, thesentence can remain indefinitely, giving the

chimp time to pick its bizarre profile andthink the matter tlhrough before responding.The permanence of the sentence not onlymakes it possible to study language without amemory problem, but to study memory in thecontext of language by regulating the durationfor which the sentence remains on the board.

In addition, because the experimenter makesthe words-the chimp merely uses them-hecan control their supply. The words availableto the chimp at any moment in time can bevaried in number, kind, type/token ratio, etc.as the experimenter chooses. The adult ani-mal, or one proficient in the language, can begiven an unlimited supply of words alongwith the opportunity to produce sentences atwill. Then the physical organization of itsvocabulary can be observed; whether, and ifso how, it lays the words out in piles to en-hance their availability for sentence construc-tion, or the degree to which it can be trainedto adopt favorable organizations. But the mainadvantages are to the training of the naive ani-mal; since the number of alternatives can becontrolled, so can the difficulty of the problem.There are no phonemes in the language, the

most elementary unit being quite deliberatelythe word. Elsewhere we lhave described a pho-nemic system suitable for the chimp (Premackand Schwartz, 1966), and can also describe asystem intermediate between the present oneand the earlier phonemic one, a system ofwords with an implicit phonemic structurethat can be made explicit at any time the ex-perimenter considers propitious. But the im-plications of these different systems are merelypractical. Although the systems vary in theirassets and liabilities, no one of them is in anysense a necessary condition for the generalfunctions of language.Human phonology is an adaptation to spe-

cific needs that are intrinsic to the human con-dition but not to language. These needs canbe eliminated, obviating the need for mecha-nisms to resolve them, and the system that re-mains is still language. For example, the prin-cipal traits of human phonology, the phonemeand the auditory channel, are both parts of asolution to a common problem. Limitationson man's memory on the one hand, and abilityto generate discriminably different responseson the other, make it impossible or highly in-efficient for him to attempt to map a largeworld by devising an irreducibly different re-

108

A FUNCTIONAL ANALYSIS OF LANGUAGE

Fig. 1. The physical basis of the language is plastic, varying in color, size and shape. Each piece of plastic isa word; the pieces are metal-backed and adhere to magnetized slate. Sentences are written on the vertical. Aword-by-word translation of the two sentences is: Sarah honey bread take; No Sarah jam cracker take. Notice theoccurrence of "no" or the negative particle as a free morph (independent word) in both Fig. 1 and 3, and inthe bottom half of Fig. 4 as a bound morph appended to "name of" forming "not name of".

sponse for each word. Rather than attempt togenerate and store 40,000 or even 5,000 differ-ent words, he produces instead only about 50or so different phonemes. By combining thesemanageably few responses, he produces thelarge number of words needed to map his com-plex world.The auditory channel also makes its contri-

bution to the same problem; indeed, in an im-portant sense it makes the combinatorial ap-proach possible. Not the auditory channel perse, but the fact that the modality of man's lan-guage is different from the primary modalityin which he perceives his world. Man is pre-dominately visual, while his language is audi-tory. It might be thought that an auditory lan-guage is merely a profound convenience-conferring the possibility of whispering whilecopulating or speaking while writing on the

board-but the contribution is of a far deepernature. To begin with, if the modalities werenot different, it would be impossible to dis-tinguish at a glance a member of the languagesystem from a member of the system referredto by the language. This simple distinction-telling a word from that which the word re-fers to-would require an inquiry of a kindthat is nearly impossible to appreciate if yourexperience with language is confined to thehuman case. You must work in an artificialsystem to appreciate this problem; in Fig. 3I have deliberately included a case where amember of the language system and a memberof the world referred to by the language notonly both belong to the same modality, butoverlap so markedly in dimensional valuesthat one can barely be told from the other.There is no sense in dwelling on this point,

109

DAVID PREMACK

however, since its impact, I fear, is very de-pendent upon experience with artificial sys-tems that do not provide for those criticalseparations so much taken for granted in thehuman system.But the minor confusions that can result

when words and designata are not immedi-ately distinguishable is the least of the dam-age that could be done the combinatorial ap-proach. If the language modality and that ofthe primary perception of the world did notdiffer, the language would very probably endup ikonic not phonemic. The tendency tomatch items on the basis of similarity is asstrong as the tendency to associate them onthe basis of contiguity, probably stronger. Butif language elements were to match the salientcharacteristics of the objects which they name,there would be as many irreducible languageelements as there are objects named in theworld. Such a development would be in com-plete opposition to the need for a small set ofmeaningless elements-meaningless in thesense that they match nothing in the world-whose combinations are used to producewords. But the only arrangement that couldfully guard against this ikonic possibility is adisparity in modality, a difference in the chan-nels in which the language is expressed andthe world is perceived. Nothing could moreeffectively preclude matching and thus allowfor the emergence of language elements fewerin number than the number of elementsnamed by the language. Thus, a primarilyvisual organism like man would require anauditory or at least a nonvisual language,even as an auditory organism if there wereone might end up with a visual language.But this whole set of problems reflects limi-

tations on human information processing; ithas little to do with the logically necessaryproperties of language. Neither the need tomap a "large" world nor a limitation in oper-ating characteristics leading to the desirabilityof a combinatorial approach to lexicon, arepressures which a system must accommodatein order to qualify as language. An artificialsubject (computer) or prosthetics added to anexperimental subject might extend memory ormake for a superior response generator; or wemight simply accept a thoroughly ikonic lan-guage: no general language function wouldappear to be precluded by an ikonic lexicon.Alternatively, we might operate upon the

world instead, reducing its elements. The lat-ter is an attractive alternative, since carryingout the basic language functions does not ap-pear to require a world of human complexity.All the essential functions of language appar-ently can be carried through in small spaces-arbitrary "corners" of a real world, or adiminished artificial world. For example, thechimp to be described here has about 40 wordsat present. In all likelihood, this could be 400,if the focus were vocabulary. But I have nointerest in vocabulary per se, not a large onein any case. Indeed, the more intriguing ques-tion is: what is the smallest lexicon in whichit is possible to carry through all the basicfunctions of language?

General Functions: WordUnlike the phoneme, which is an adjust-

ment to a nonessential problem, the word isan essential unit of language. It reflects someof the most basic features of experience, amongthem the consensus that perceptual experiencecan be divided into stable elements, and theagreement that all such elements can be repre-sented or referred to by responses of the or-ganism.The introduction of language to a totally

naive organism is substantially different fromintroducing new words to an organism alreadyequipped with a bit of language. "X is thename of Y" is a powerful device that can beused to teach words to the advanced subject,as is shown in the section on metalinguistics.But here we are talking of first words and atotally naive subject. With this subject, a vitalfirst step is to establish a simple social trans-action between the subject and the trainer. Weestablish this transaction well, assuring our-selves that it holds no concepts of which thesubject is incapable, for it is this transactionwe will map with language. Mapping, as themany examples to follow will show, amountsto dividing a routine into its componentclasses, displaying the range of values thateach class can take, and assigning a name toeach of the values.A feeding routine makes an effective trans-

action with a young chimp. The trainer placesan edible item on the table between him andthe subject and looks by benevolently whilethe chimp takes it and eats it. Once or twice,after giving the fruit to the chimp, the trainermay "request" that the fruit be given back to

110

A FUNCTIONAL ANALYSIS OF LANGUAGE

him; he extends a cupped hand and extrudeshis lip to the best of his ability, mimicking thesupplication chimps direct at one another, butthe imitation is unsuccessful. The chimp lookspuzzled, even hesitant, but never gives backthe frUlit, so the trainer goes on with that partof the routine that is successful, laying fruitout and watching the chimp eat it.Then one day, after the transaction is well

established, the trainer places an element fromthe language system, a piece of colored plas-tic, alongside a piece of fruit, say, a banana.The banana is now farther back than usual,out of reach, while the plastic chip is forward,easily within reach. The animal is induced tomake a prescribed response with the languageelement, in this case, place it on the languageboard, after which she is given the fruit. Thechimp is almost immediately proficient in thisact. Causing objects to adhere to a verticalsurface is something it does readily in con-trast, for example, to producing human sound.(Notice that in this system, unlike the humanone, production need not lag behind compre-hension. The subject does not make butmerely uses the words, and can do so fromthe beginning without having to undergoelaborate motor learning. Thus, the earliesttraining can occur in the production, as wellas in the comprehension, mode. This can bean advantage in training young subjects, sincethe control of attention is more certain whenthe subject is required to respond rather thanmerely observe.)What does the piece of plastic mean to the

chimp? The question is hopelessly prematureat this stage for, in a sense, all of the trainingthat lies ahead is an attempt to inculcate asystem that will make it possible to answerquestions of that kind. The question is one towhich the subject's own answers can con-tribute importantly, provided we can makethe subject susceptible to questions. How toconfer question-answering ability upon achimp is illustrated in the section on theinterrogative.The rest of the training consists of making

simultaneous changes in some aspect of thetransaction and in some aspect of the languagesystem, so as to establish a correspondencebetween the two systems. For example, wemay start with the fruits that are offered. Theset of possible objects is defined by offeringdifferent fruits on different trials and each

time with a corresponding change in the lan-guage element. When the fruit is banana theplastic chip is of one kind, when apple of adifferent kind, and when orange still a thirdkind. On each trial the chimp's task is thesame: place the piece of plastic that is along-side the fruit on the board before receivingthe fruit.Two kinds of tests will show whether or not

the subject has formed an association betweenmembers of the object class and of the cor-responding language class. Trials on whichthe chimp is given two would-be words butonly one piece of fruit will determine whetherit can match the word with the fruit. But thesubject may know more than such choicetrials reveal. If the subject is less interested inthe fruit that is offered on a particular trialthan the fruit that is not offered, it may usethe "wrong" word essentially as a request -forthe fruit that it prefers. This possibility canbe detected by obtaining independent prefer-ence orderings on the fruits and on the words.For example, allow the subject to choose be-tween all possible pairs of fruits and on an-other occasion all possible pairs of words. Ifits preferences among the words agrees withits preferences among the fruits, then the sub-ject must know what word goes with whatfruit, whatever its choice behavior may sug-gest to the contrary.The next perceptual class is mapped in the

same fashion. Each change in fruit was accom-panied by a change in the language element;now, in similar fashion, each change in donor-the person giving the fruit-is accompaniedby a change in the second language element.For example, when Mary is present and thefruit is apple, the chimp (Sarah) must write,"Mary apple," to receive the apple; withRandy present, "Randy apple," etc. Associa-tions for members of the donor class can betested in the same way as for members of theobject class. With one trainer present but twoor more donor words, the subject must matchthe word to the trainer. Similarly, if necessary,a preference ordering can be determined fordonors' and would-be donors' names, and theconcordance between the two orderings takenas evidence of an association.

In addition to being required to place twopieces of plastic on the board, the chimp isrequired to observe a proper order. "MaryApple" is accepted but "apple Mary" is not.

III

DAVID PREMACK

In the sentence toward which she is progress-ing, words will occur in the one order but notin the other. Thus, in the target sentence,"Mary give apple Sarah" we find "Mary apple"but not the reverse. The correct order is re-quired from the beginning so that incorrectorders will not have to be unlearned later. Inaddition to the order requirement, we observean anti-regression rule. Once she reaches atwo-word state, we reject one-word fragments,even as we reject two-word fragments when shereaches a three-word stage, etc. (This rule isrigidly enforced only at this tender stage,when there is uncertainty not only in her useof language but, more important, in our esti-mate of how much language she actuallyknows. At a later stage, fragments are wel-comed, it being of interest to see what the pro-portionality may be between her use of frag-ments and the redundancy in the situation.)The fruit and the donor are easily mapped;

the other two classes in this example presentdifficulty, though of a practical kind. For ex-ample, the attempt to map the recipient byvarying who it is that receives the fruit runsinto difficulty of a predictable kind; the chimpis reluctant to produce response sequencescalling for a recipient other than herself. Simi-larly, the attempt to map the operator, byvarying the action upon the fruit-sometimesgiving it as before but other times cutting itor inserting it in a pail-encounters the sameproblem; some of the outcomes are so non-preferred that, once she has associated the lan-guage element with the undesired action, shewill not form the sentence. These are strictlypractical problems, however; they can bedealt with for the most part by arranging ap-propriate contingencies; e.g., when Sarahwrites "Mary give apple Jim," thereby effec-tively denying herself the apple, she can begiven a tidbit more preferred than apple. Ifreinforced, altruism can become quite reliable.The order in which the transaction is

mapped-fruit, donor, operator; or donor, re-cipient, etc.-is not something we have testedyet. Also, it may make a difference how manymembers of one class are established before anew class is introduced; this too remains tobe tested. Notice, however, that while we maypuzzle these minor parameters, we have nocomparable hesitation in deciding a far morebasic matter, viz., how to partition the trans-action in the first place. Why is this so?

The effect of the order of the mapping canreadily be converted into an experimentalquestion, as can the number of entries thatshould be established in one class before pro-ceeding to another, for we have no difficultyin proposing alternative orders or numbers.But the effect of alternative partitionings, ofdividing the transaction one way rather thananother, does not go over into an experimentalquestion, for we are incapable of proposingsignificantly different partitionings. We see thesituation one way and one way only. There isa recipient, a donor, the object that is ex-changed, and the act of giving. All alternativepartitionings that we may propose will turnout to be trivial variations on this one, pro-posals either to omit a class or to slice anexisting class more finely. Our freedom in thismatter consists in our choice of situations tomap; thus, rather than start with feeding, wemight begin with body care and lead into theoperator "wash" and names for parts of thebody. We are not free in how to divide thissituation. We are free to choose the situationbut not free in our choice of how to divide it.These perceptual constraints must have a

great practical advantage, however. The or-ganism being trained in language could notbe less circumscribed perceptually than weare. If both organisms see the situation thesame way this must greatly facilitate the train-ing, in ways that we have not yet fully workedout.

In summary, while many variations remainto be explored, the basic procedure for teach-ing words to a naive organism is extremely sim-ple. A transaction is established between thesubject and the trainer and a decision is madeas to the salient perceptual classes into whichthe transaction should be divided, a decisionthat will prove to be remarkably easy. Theneach class is rotated through a series of values,in the present case, apple, banana, etc.; Mary,Randy, etc.; give, insert, etc.; Sarah, Jim, etc.,the other classes being held constant. As thevalue of the perceptual class is changed, acorresponding change is made in the languageelement. And as each new class is mapped,the language requirement is increased. In thebeginning, the chimp merely took the fruit,looked over benevolently by the trainer. Thenone word was required, two, three, until finallytarget sentences were realized such as "Marygive apple Sarah."

112

A FUNCTIONAL ANALYSIS OF LANGUAGE

General Functions: SentenceThe subject may produce properly ordered

strings of words and yet give no essential evi-dence that the string is a sentence. A sentencediffers from a string of words in that it hasan internal organization, a knowledge of whichis a necessary condition for correctly respond-ing to it. The knowledge can be representedby a tree diagram of the sentence or the appli-cation of parentheses, both of which will showthe dependencies between any one word inthe sentence and all the others (e.g. Chomsky,1965). I will deal with this topic under twoheadings-two-term relations and hierarchicalorganization-in the course of which we willsee some examples of the dependencies thatdistinguish a sentence from a string of words.

Symmetrical RelationsThe conditions leading to the necessity for

syntax are general, involving information thatis commonplace rather than exotic. Perhapsthe simplest condition is the symmetrical two-term relation such as can be found in geo-

metrical propositions, for example, and insome of the verbs of social behavior. Thesesemantically disparate topics have in commonrelations whose terms can be interchanged.For example, A on C but also C on A; X talksto Y but also Y talks to X. This interchange-ability makes it impossible to distinguish phy-sically the class of items that can take one posi-tion in the relation from the class that can

take the other position. In the limiting casethe membership of the two classes can be com-

pletely interchanged, so that, for example, any-

thing that can be on top can also be on thebottom, and anything that can talk can alsobe talked to. In cases of this kind there is noother way to identify the position of the itemin the relation than by the order of the cor-

responding word in the sentence. This stateof affairs contrasts with what we may call a

closed relation, where there is a well-markeddifference between items that take one positionin the relation and those that take the otherposition, and for which a semantic covariationrule is a sufficient kind of organization.

For example, the verb "insert" as we haveused it with Sarah involved a closed relationand could be dealt with by a semantic rule.In Sarah's experience as in ours, she (andMary and Randy and Jim) insert(s) pieces of

fruit in dishes and pails, but pieces of fruitdo not insert her (or Mary or Randy or Jim).Inserters are one kind of thing, insertables adifferent kind. "Insert" could therefore betreated as a relation between two classes de-fined on the basis of physical or functionalproperties of the membership, which is whata semantic covariation rule amounts to. Onthis treatment "Sarah insert banana," "bananainsert Sarah," "insert banana Sarah," etc.should not differ from one another; if the sub-ject's training included experience with vari-able word order, so as to safeguard emotionalreactions to novel forms, its response shouldbe the same to all possible forms of the string.(Notice that one limitation of semantic co-variation rules is already evident. The worldneed not change, and oranges begin to insertMarys, in order that we may want to talkabout such possibilities. But such talk wouldnot be possible with semantic covariationrules; semantic rules set limits on speculativediscourse that syntactic formulations do not.)The prepositions "on", "under", and "to

the side of" involve physical relations whichwe do not doubt that the chimp can discrimi-nate and which can be trained in a mannerthat will assure a complete interchange of theitems that take the two positions in the rela-tion. To evaluate the chimp's capacity for syn-tax we must make certain that the preposi-tions are in fact learned on a syntactic basis.This is possible only if the application of thepreposition is restricted to so-to-speak seman-tically neutral domains-cases where one formof a relation is as sensible as the other. Forexample, objects that differ only in their colorwill fulfill this condition, for red on greenhas no semantic edge over green on red,unlike, for example, fly on horse, which is no-tably more probable than the reverse. This re-striction is necessary because, although prepo-sitions are in principle symmetrical, they maybe used asymmetrically in the beginning anddefined improperly as a result. For instance,the regularity with which children go onbikes, cups on saucers, plates on tables, lampson floors, etc. may lead the child to define "on"as a semantic rule in which the larger of twoitems goes on top and the smaller on thebottom. Only accidents in which, for example,bike ends up on child or plate ends up onfood, not to mention later sexual experience,may induce the child to abandon the initial

113

DAVID PREMACK

semantic formulation and redefine the wordon purely syntactic grounds. Since, however,our control of the training sample is betterthan that of the average parent, we need notrely on accidents to assure that our subjectuses syntactic definitions from the beginning.Four color words-"red", "green", "blue",

and "yellow"-had been taught the subjectearlier, and we used them in training theprepositions. Two by four cards, painted oneof the four colors, but indistinguishable other-wise, were used as the objects. The cards wereplaced on top of each other with the top oneoffset a bit so that the bottom one could beseen. The four color words make possible 12different cases, e.g., "red on green", "green onred", "blue on red", etc. "On", "under", and"to the side of" have special interest in a lan-guage written on the vertical. The order of theelements in the sentence and the order of theelements referred to by the sentence are thesame in the case of "on", opposite in the caseof "under", and unrelated in the case of "tothe side of". We are interested in determiningwhether the isomorphism between the twoorders or the lack of it will affect the rate oflearning; unfortunately, at the moment, "on"is the only one for which the training has beencompleted.

Using the colored cards as described, wetrained her in the comprehension mode, whereshe was required to respond to our sentencesand then subsequently tested her ability toproduce the same sentences herself. The train-ing proceeded in three steps, the first restrictedto one pair of colors, the second dealing withher ability to generalize to new colors, and thelast examining her transfer from comprehen-sion to production.

In the first step of the training, the red cardwas placed on the table before the subject,the trainer wrote on the board, "green on red",handed Sarah the green card and then inducedher to place it on the one that was alreadythere. I describe this procedure, which is in noway unusual, simply to illustrate the generalstrategy of a strict training procedure: tobring about the desired behavior by limitingthe possibility for other kinds of behavior. Thestep used is hardly the only possible one, andhad it failed we would have tried others,which illustrates a second characteristic ofstrict training procedures: they are based onjudgment not algorithms.

Next, the opposite sentence "red on green"was presented, now with the green card downand the red card handed to the subject as theone to be placed on top. Subsequently, thesubject was given both cards and presentedfirst with one form of the sentence and thenwith the other. Once she was proficient at pro-ducing the card arrangement called for by thesentence, she was given sentences using all fourcolor words, e.g., "yellow on blue", "red onyellow", etc. She performed as well on the 10new cases as on the two training cases (red,green), demonstrating that she had not simplymemorized the training cases but could applythe preposition to new cases. (These tests donot demonstrate that she could use "on" inany domain other than color-or indeed evenwith colored blocks rather than cards-but thatquestion, though intriguing, is not germaneto the present discussion of syntax.)The last step concerned her ability to pro-

duce sentences appropriate to the trainer's be-havior, rather than, as in the initial training,to behave in ways appropriate to the trainer'ssentences. On each trial she was given threewords, two color words and "on", and requiredto place them on the board in a way thatcorresponded to, or described, the trainer'splacement of the cards. Thus, if the trainerput the blue card on the green one, Sarah,who held the words, "green", "blue", and "on"in her substantial hand, was required to write"blue on green" and not vice versa. Her pro-ficiency at this was 80% in the first set of 10trials, which is indicative of a high order oftransfer, since about 80% correct is her usualasymptotic performance level for essentiallyall problcms.

Hierarchical OrganizationConsider the sentence, in chimp language,

"Sarah insert banana pail apple dish." Trans-lated into English this is the instruction toseparate the banana and apple, to put the for-mer in the pail and the latter in the dish. Butto carry out that quite simple instruction, orinstructions of that general kind, requires aknowledge of the internal organization of thesentence. That organization can be shown byusing parentheses to indicate the dependenciesamong the words. For example, (Sarah insert[(banana pail) (apple' dish)]) shows that "ba-nana" and "pail" go together, likewise "apple"and "dish"; that "insert" applies not only to

114

A FUNCTIONAL ANALYSIS OF LANGUAGE

"banana-pail" but to both cases, and finallythat it is Sarah who is to carry out the action.This example carries us beyond word order

to a second contribution of syntax-hierarchi-cal organization of the sentence. How essentialis that factor? What would be lost if, for ex-ample, the strings of words in the languagewere sensitive to order but were not organizedhierarchically? We can answer that questionby comparing the subject's behavior to theabove sentence under conditions in which anunderstanding of it was based on either ofthree levels of organization: (i) word knowl-edge or gross semantic rule, (ii) refined seman-tic rule, or (iii) hierarchical.

Several years ago in working with psychoticchildren, R. Metz and I devised some tests oflanguage comprehension to determine whetherthe severely impaired speech production,which characterized these children, was owedto performance factors or to something deeper.Having earlier confirmed Lenneberg's (1967)surprising claim that the feeble-minded childis grammatical, I was distinctly surprised whensome of the psychotic children failed all ofthe comprehension tests (while at the sametime performing adequately on nonlanguagetests). They had what we ended up calling"word knowledge" but, so far as our testscould determine, little else. Response to thesentence in question at the level of wordknowledge would amount to the following.Understanding of the word "Sarah" would

result in Sarah's carrying out whatever actionwas carried out rather than waiting for, say,Mary to act, as she would on the occasion ofsentences that began with "Mary". Secondly,an understanding of the word "insert" wouldassure action of one kind-putting one thinginto another-as opposed to cutting, taking,giving, some of the other verbs or operatorsthat she knows. Third, the objects acted uponwould be confined to those named. But withno more than word knowledge, there would begreat latitude in what was inserted into what.Every object could be inserted into everyother; thus, the dish could go into the pailas readily as the banana go into the dish, etc.Clearly, the specific pairing of banana withpail and apple with dish that would aloneconstitute evidence of an understanding of the(hierarchically organized) sentence would notbe guaranteed by a knowledge limited to indi-vidual words.

A higher level of organization could be pro-vided by semantic covariation rules; theycould limit the possible outcomes, bringingthem closer to the desired one, but would stillfall short. Actually, what we have called wordknowledge could be formulated as a coarsesemantic covariation rule; the second level oforganization would then amount to the addi-tion of a second semantic rule which wouldhave the effect of refining the first rule. Thus,the first case could be analyzed in a manneralready suggested above, as a rule in whichthe verb was the predicate and the agents andobjects were the arguments taken by the predi-cate. This could be diagrammed as:

insert

agents objects

emphasizing the focal role of the verb or op-erator and the secondary role of that whichinstanced the verb, inserters on the one hand,and objects of insertion on the other.The addition of a second semantic rule

could differentiate the object class, separatingthe containers (dish, pail) from that which getscontained (apple, banana, etc.). This betterdefined situation could be diagrammed as:

insert

agents objects

fruits containers.

The additional definition of the semanticspace would further delimit the subject's be-havior, e.g., it would keep the dish out of thepail, but it would no more assure the exactoutcome indicative of an understanding of thehierarchically organized sentence than the firstorganization. Indeed, no combination of se-mantic rules could assure that outcome since,so to speak, pails and dishes will accept onekind of fruit as readily as the other.

Training Procedure for theCompound Sentence

I will now describe the procedure we usedto teach the chimp to respond correctly to thesentence in question, though without claimingthat she is therefore unquestionably proficientin syntax. Proving competence in syntax isticklish, not impossible, but difficult or atleast arduous. Testing any highly inferential

115

DAVID PREMACK

matter is, of course, more troublesome thanwhen the hypotheses lie closer to the surfaceof the data. The problem is aggravated by afactor which the reader might guess simply byrecalling his own childhood. Grammar is notthe chimp's favorite subject. There is a limitto the number of tests she will accept on agrammatical topic, and that limit is not alwayssufficient to include all the control sentencesone might desire.The training proceeded by three steps, all in

the comprehension mode. First, she wastrained individually on each of the four sim-ple sentences of which the compound sentencein question could be composed. These in-cluded:

Sarah insert banana pail.Sarah insert apple pail.Sarah insert banana dish.Sarah insert apple dish.

The trainer wrote each sentence on the board,at the same time offering a choice of fruitsand containers, and requiring the subject toplace the designated fruit in the designatedcontainer. Next, she was given all possiblepairs of the sentences, side by side, in the man-ner of a paragraph. For example,

Sarahinsertbananadish

Sarahinsertapplepail.

Since no change was made in the compositionof the individual sentences, this step was in-tended merely to accustom her to carrying outtwo acts of insertion, as is required by thecompound sentence.

In the final step, all possible pairs of sen-tences were again combined, this time one

immediately above the other, and this conjunc-tion of two simple sentences was graduallyconverted into one compound sentence. Theprocedure generated sentences of this kind:

(I)

SarahinsertbananapailSarahinsertappledish

(ii)Sarahinsertbananapailinsertappledish

(iii)Sarahinsertbananapailappledish

Neither the deletion of the second use of"Sarah" nor the subsequent deletion of thesecond use of "insert" disrupted her perform-ance. She was performing at between 75 and80% correct, her customary level, and con-tinued to do so when the changes were made.Nor was the subject disrupted by the subse-quent generalization tests we have carried out.These tests are incomplete in that while wehave substituted fruit names (e.g., grape, apri-cot, orange for those used in training, we havenot yet changed the verbs, and these may pre-sent more difficulty. Also, there are some adhoc rules that she might be using in processingthe compound sentences, not all of which havebeen explicitly eliminated by test. I will de-scribe just one as an example of the severalthat could apply.

For example, she might use a rule that said,in effect, apply the container word to the ob-ject word above it. This would work for theindividual sentences but would run into trou-ble when one sentence was written above theother; then the bottom-most container wordwould apply to two object words, one directlyand one indirectly above it. In the compoundsentence in question, the rule would lead herto apply dish correctly, but pail incorrectly,for pail would apply to both apple and ba-nana both of which are above it. There wasno evidence for an error of this kind.This error could be overcome, however, if

the rule were expanded to read apply thecontainer word to the fruit word immediatelyabove it. But this modification would fail inthe case of sentences of this kind:

Sarahinsertappledishapplebananapail

for apple would go into dish but would failto make it into pail. That is, the effect of"immediately" would be to confine pail tobanana. This is one of the control cases wehave not yet tested.

Notice that these ad hoc rules, though notyet explicitly eliminated by test, are not neces-sarily more parsimonious than the kind ofrules that could generate the hierarchical or-ganization which could equally well account

116

A FUNCTIONAL ANALYSIS OF LANGUAGE

for her present performance. Nevertheless, thequestion of her syntactic competence must re-main undecided for the time being, until wehave examined a larger portion of the sen-tences that she is capable of making.

Interrogative and the Concept ofSameness or Identity

Except for formulations like "red on green,"etc., which the chimp produced in conformityto the trainer's placement of the cards, andwhich can be interpreted as declaratives, theother sentences-both those of the trainer andthe subject-were in the imperative mood;e.g., "Mary give banana Sarah," "Sarah takebanana pail orange dish," etc. There havebeen no questions and no metalinguisticpropositions. By the latter I have in mind theuse of language to teach language, as in thecase of "X is the name of Y," where a so-far un-named object is given a name, or "X is a syno-nym for Y," where an object with one nameis given another one.We will treat the interrogative first. It can

be introduced in several ways, of course,though we found highly effective a procedurethat exploited the concept of sameness oridentity. Thus, some of the very first questionswe asked the chimp were, in English para-phrase, "What is the relation between thoseobjects, identity or non-identity?" Identity orsameness is a concept for which the chimp canreadily distinguish between positive and nega-tive instances. It is therefore an ideal subjectmatter for questions, since in the beginning atleast there is no point in asking the chimpquestions she cannot answer. Also, the pro-cedure leads naturally from the wh- form ofthe question to the yes/no form. Hence, theadvantage of the procedure is that the chimpis asked questions she can answer and theseveral forms of the question develop in anatural succession from a common situation.What are the concepts that the chimp

knows? Fortunately, we need not guess butcan answer that question directly throughmatching-to-sample procedures. For example,if given a cup and a spoon, and another cupset somewhat apart from the other two ob-jects, the subject will indicate by whatevergesture we choose, that the cups go togetheror participate in a relation which the cup andspoon do not. This procedure, which does notdepend upon language and which can be ap-

plied broadly, is a boon to the teaching oflanguage. With it we can determine what it isthat the subject already knows and thus canavoid wasting time attempting to teach thesubject names for concepts it does not know.Only with rare exceptions-of a kind I willdeal with in the section on metalinguistics-does language teach the subject new concepts.Mostly it merely teaches names for conceptsthe subject already knows.The matching procedure is not restricted to

object identity but can be used to inventorythe overall conceptual knowledge of the sub-ject. For example, the subject can be given ob-jects that match on no other basis than that ofcolor, or shape, or size; or sets of objects thatmatch on no other basis than that of area,volume or number. Similarly, it can be usedto assess the capacities for perceptual trans-formations. The subject can be required tomatch (three-dimensional) objects to their(two-dimensional pictures), or vice versa-something that Sarah does nicely-as well asmatch one orientation of an object or pictureto another orientation. Whenever it is shownthat the subject can distinguish between posi-tive and negative instances of the concept inquestion, it is reasonable to attempt to intro-duce a name for the concept. For the pro-cedure by which a name is taught for the mostabstract concept is no different from the oneby which a name is taught for the most mun-dane object.Having established that the subject is capa-

ble of matching like objects, a name is estab-lished for the fact of sameness or identity-aswell as a name for non-identity-in the sameway names were previously established for ob-jects, actions and agents. We establish a con-sistent relation between positive instances ofthe concept and one language element andnegative instances and a second language ele-ment. In the actual training, two objects, saytwo cups, were placed before her at a smalldistance from one another. She was given apiece of plastic, intended to mean same oridentical, and required to place it between thetwo like objects. On other trials she was givenobjects that were not the same and requiredto place a different language element betweenthem, one intended to mean non-identical ordifferent. (Notice how an exactly analogousprocedure can be used to map the concept ofequality and inequality. For example, three

117

DAVID PREMACK

marbles are set slightly apart from three sticks,and the subject is required to place an equalitysign between them. On other trials, two mar-bles are set apart from five sticks, and the sub-ject is required to place an inequality sign be-tween them. Both the schema for identity andfor equality are so far hybrid, i.e., contain ele-ments from the world and from the systemmapping the world. For example, XXX =YYY, or stick, stick, stick = marble, marble,marble, rather than 3 = 3. To convert the hy-brid sentence into a pure form requires nam-ing the objects so that the relation can bestated between names of objects rather thanobjects themselves. Likewise, converting thehybrid equation into a pure form requiresnaming the number of the objects so that therelation can be stated between numbers ratherthan number of objects. The traditional ques-tion about the arithmetic ability of infra-human primates has been misformulated. Canchimps be taught to count or to do arithmetic?That is a misleading way to put the question.A better way is: do chimps recognize numer-osity? That is, can they match sets of objectsthat have nothing in common except the factof their equal number? That question can beanswered, in a matter of hours, with matching-to-sample procedures. If the answer is positive,there would be little doubt but that theycould be taught to count. But if the answer isnegative and the chimp cannot recognize nu-merosity, though some might remain sanguineI would not wish to be the instructor. Teach-ing a concept and mapping one that alreadyexists are different enterprises.)Were the proper associations formed be-

tween the language elements on the one handand cases of identity and non-identity on theother? That was established in the usual man-ner. She was given both the words for sameand different in the presence of a case of, say,identity on one trial and non-identity on thenext, and required to match the word to thecase. She was also given generalization tests todetermine whether she had merely memorizedthe training cases. We tried her on a variety ofnew cases, including new combinations of olditems, entirely new items, and new items forwhich names had and had not been previouslytaught. But these experimental niceties provedto be in vain, for she made so few overall errorsthat one condition could not be differentiatedfrom another. Also, she made no more errors

on non-identity than on identity. Thus, inprinciple, the subject could go about the cage,picking up pairs of objects, and labeling themidentical or non-identical. Any instance of thiskind, which she could previously recognize,she could now label as such. That is what thelanguage training contributed. The matching-to-sample procedure established that she coulddistinguish between positive and negative in-stances of identity; the language training sim-ply provided her with a name for a conceptthat already existed.

InterrogativeIn the identity exercise above the chimp is

already being asked a question. If you wereto instruct an English-speaking subject in thesame task as the chimp's, you would say some-thing along these lines:"What is the relationbetween the two objects, are they the same ordifferent?" The chimp is asked the same ques-tion but witlhout benefit of an explicit inter-rogative marker. The only marker she has sofar is the implicit one of the space betweenthe objects-into wlhich she is to insert heranswer-along with the fact that the trial doesnot end until she completes the schema byadding the third item.The question can be made explicit by any

one of the three standard linguistic devices:inflection, word order, or an interrogative par-ticle. We chose the latter as the simplest bothin the sense of involving the least change forthe subject and of being most compatible withthe present physical system. So we simplyadded a question marker to the schema shewas already receiving. For example, where wepreviously wrote:

A A"identical" "non-identical"

we henceforth wrote

A ? A"identical" "non-identical"

We are now in a position to look at exam-ples of all there types of questions which theidentity exercise generates. The first two arewh- types-what or which-while the last oneis of the yes/no form. Examples of two ver-sions of a wh- type question are shown in theupper panel of Fig. 2; they can be paraphrasedas "X is what to X?" and "X is what to Y?".The alternatives for both versions are "identi-

118

A FUNCTIONAL ANALYSIS OF LANGUAGE

Fig. 2. Four Wh-type questions, with English paraphrases.

cal" or "non-identical" and the subject's taskin both cases is to replace the interrogativemarker with the word appropriate to the case.

Two versions of a second type of wh- ques-tion are shown in the lower panel of Fig. 2;they can be paraphrased as "X is the same as

what?" and "X is not the same as what?" Thealternatives are no longer the words "identi-cal" or "non-identical" but the objects them-selves. The subject's task remains essentiallythe same, however: to replace the interroga-tive particle with the appropriate object.The yes/no question, which is shown in

Fig. 3, has four forms rather than two as inthe wh- questions, and these can be para-

phrased as (i) "Is X the same as X?", (ii) "IsX not the same as X?", (iii) "Is X not thesame as Y?", and (iv) "Is X the same as Y?".Her alternatives now are neither the words"identical" or "non-identical" nor the objectX and Y, but the words "yes" and "no". Her

task remains the same: to replace the questionmark with either "yes" or "no".The chimp is capable of answering correctly

all the question types shown, for an essentiallyunlimited variety of items, words as well asobjects. In this case, too, the generalizationtests, requiring that she recognize the conceptwhen applied to items not used in training,proved that she had not merely memorizedthe training cases but could apply the conceptbroadly. Indeed, the subject has never faileda generalization test. Though often trained onno more than two positive and two negativeinstances of the concept, she has always beenable to apply the words to cases not used intraining. How shall we explain this? If weadopt the position that in teaching a new wordwe are at the same time teaching a new con-cept, generalization emerges as a mystery. Butif we recognize that concepts antedate the lan-guage training, there is then little mystery in

119

DAVID PREMACK

Fig. 3. Four yes/no type questions, with English paraphrases. Notice how one of the objects shares dimen-sional values with the language elements and, though not a word, is easily confused with the words.

the subject's ability to apply words beyondthe training cases. Only failures in generaliza-tion would speak against the existence of con-cepts that antedated the language training.

Consider three objections to calling theabove expressions questions. First, each casehas the form of a one-to-one substitution: Theinterrogative particle is removed and replacedby a single item. Human interrogation doesnot suffer from this limitation since answersmay be of any length.2 But in fact, the presentsystem is not restricted to one-word answerseither. The answers considered have beenshort because the language is simple; in addi-tion, there are already counter-examples inwhich answers consist of more than one word.For example, if we give her as alternatives,"no" and "identical"-rather than "identical"

2I am indebted to E. Klima for this interesting ob-servation.

and "non-identical"-when the question askedof her is "A ? B," i.e., what is the relation be-tween A and B?, she will answer "no identi-cal," which though laconic is nonetheless atwo-word answer and is thus no longer a caseof one-to-one substitution. (Her ability to usethe negative particle in this fashion is of con-siderable interest in its own right, though notto the present discussion.)

Second, she has answered but never askedany questions. This would be a serious objec-tion if she failed to ask questions when givenan opportunity to do so. But so far the omis-sion is in the training program, not the sub-ject. Our failure to have tested her ability toproduce questions is a failure to have found asimple condition in which to make the test.(On one occasion the chimp offered her ownsolution to this problem. She stole the testmaterials for the lesson, as she does from timeto time, and went on both first to produce and

120

A FUNCTIONAL ANALYSIS OF LANGUAGE

then to answer many of the questions we hadtaught her.)

Third, a quite different kind of objectionmight be made on the grounds that the ques-tions are not themselves genuine. In humanaffairs, a question is often taken as evidence of(i) the speaker's ignorance and (ii) his assump-tion that the listener is less ignorant. Thus, aperson may ask, "What time is it?" because (i)he does not know and (ii) he believes hislistener does know. It is clear that the ques-tions asked the chimp do not belong to thisparadigm.But it is equally clear what paradigm they

do fit and also that the paradigm above is onlyone of several that figure in human interroga-tive behavior. Another strong paradigm isbased on exactly the opposite assumptionabout the distribution of knowledge; it is themainstay of the teacher's business and is theparadigm that is applicable here. The teacherknows the answer to the question and there-fore asks it of the student to find out whetherhe too knows the answer. Thus, a father mayask his son, "What time is it?" thereby deter-mining whether or not the boy can tell time.Although the distribution of knowledge is nowthe opposite of that in the first case, the utter-ance is no less a question.

MetalinguisticsThe relation that now concerns us can be

paraphrased as "X is the name of Y." In therelation just dealt with, viz., "X is identicalwith X," the items instancing the relation wereboth objects. But the items instancing thepresent relation are a word and an object, aword which in positive instances of the con-cept is the name of the object, and a wordwhich in negative instances is not the nameof the object. The training procedure for es-tablishing a name for this concept is nonethe-less the same as for all the cases alreadydescribed. In brief, positive instances of theconcept are associated with one language ele-ment, negative instances with another.The well-established words for apple and

banana along with the corresponding objectswere the training materials used. Thus, posi-tive instances consisted of the pairs, "apple"-apple and "banana"-banana, and negativeinstances of the pairs, "apple"-banana and"banana"-apple. That is, positive instancesconsisted of a word and the object named by

the word, negative instances of a word and anobject not named by the word.The subject was given a word-object pair

set slightly apart and required to place be-tween them the language element intended tomean "is the name of". Parallel trials weregiven on negative instances, i.e., a pair such as"banana"-apple was set before her and shewas required to place between them the lan-guage element intended to mean "is not thename of". Should we proliferate the vocabu-lary by introducing an independent word forthe negative case, or should we economize andrequire her to form the name for the negative("is not name of") by applying the negativeparticles to the name for the positive case? Wecompromised by introducing an independentname for the negative case which consisted,however, of the negative particle attached as asingle unit to the name for the positive case(see Fig. 4).

Following the training in which the words"name of" and "not-name of" were associatedwith the negative and positive cases respec-tively, her associations were tested in a mannerthat took advantage of her competence in theinterrogative. First, she was asked a wh- typequestion, in effect, "What is the relation be-tween 'apple' and apple?" Her alternativeswere "name of" and not-name of" and her taskwas to replace the interrogative marker withthe word of her choosing. Next, her associa-tions were tested with yes/no questions. Forexample, she was asked in effect, "'banana' isthe name of apple?" and required to choosebetween the alternatives "yes" and "no". Shewas tested on both versions of the wh- ques-tion as shown in Fig. 2, and all four versionsof the yes/no question as shown in Fig. 3. Thequestions were confined at first to the fourword-object pairs on 'which she had beentrained, but once it was clear that she hadmastered the training cases, this restriction waslifted and the same kinds of questions wereasked of a number of word-object pairs thathad not been used in training, some of themthe names of fruits like those used in trainingand some of them not, e.g., "dish" and "pail".Once again her performance on the generali-zation test did not differ materially from herperformance on the training materials, indi-cating that she was able to apply the namesbeyond those few cases on which she had beentrained.

121

DA VID PREMA4CK

Fig. 4. Examples of the relation "X is the name ofY" and "X is not the name of Y". Note that the name

for the negative relation amounts to the negative par-ticle built onto the name for the positive relation.

Productive vs. RecognitionalUse of a Concept

I have emphasized 41 along that languagetraining is primarily a mapping of existingknowledge; it gives the subject names for dis-tinctions that he or she could draw in advanceof the language training. In the concept"name of" we have an important exception.This concept was inculcated by the languagetraining and would not have existed otherwise.An interesting and powerful consequence ofthis fact is that this kind of concept can beused to generate new instances of itself.

In the standard generalization test, the sub-ject may recognize instances of the concept thatwere not used in training, but this is not the

same thing as using a concept to generatenew instances of itself. In the generalizationtest, the cases called "new" are so only in thelimited sense that they were not used in thetraining of the word in question. They areotherwise old, established cases. For example,when we found that the chimp could apply"name of" to "apricot"-apricot (and "not-nameof" to "apricot"-raisin), we proved that shecould apply the words to -cases other thanthose used in training. But the pair "apricot"-apricot was not a new pair. It was an old pairwitlh a history of use in sentences that thechimp had both produced and comprehended,some part of which history-exactly which partwe do not yet know-enabled the chimp toidentify the pair as an instance of "name of".This is quite different from introducing newwords by telling the subject that this piece ofplastic is the name of this object; and thenfinding that the piece of plastic is henceforthused appropriately in sentences dealing withthe object in question. We will call the appli-cation of a word to new cases in a generaliza-tion test recognitional, to distinguish it fromthis second and more powerful use which wewill call productive.

Fig and crackerjacks were objects that inter-ested the subject but which had never beengiven names. We used them as the first objectsto name with the use of the concept "name of"rather than the more laborious procedure ofrotating a class through a range of valueswhile at the same time changing the corre-sponding language elements (which, however,is the only procedure that can be used to teacha naive subject its first words. See above for adescription of this more laborious procedure).A three-step procedure was used to introducethe new words, as well as test the effectivenessof the introduction by subsequently requiringher to use the word in a sentence.

First, a piece of plastic (potential word) anda fig were placed slightly apart and the word"'name of" set between them; next a secondpiece of plastic was placed slightly apart froma fig and the word "not-name of" was placedbetween them. Had the subject paid attentionto the lesson or had her attention wanderedlike that of students in classrooms everywhere?In order to answer this question we resortedto the interrogative, as teachers do everywhere,giving the subject both the wh- and yes/noforms of the question. In effect, she was asked,

122

A FUNCTIONAL ANALYSIS OF LANGUAGE

"What is the relation between the piece ofplastic and fig?" for which the alternativeswere "name of" or "not-name of" and "Is thispiece of plastic the name of fig?" where thealternatives were "yes" and "no". Her gener-ally correct answers permit us to move to thelast step, which required that she use the ap-propriate word in a sentence.The materials set before the subject were a

fig and a number of words; "fig", the piece ofplastic she had been told was not the name offig, the names of two other fruits, "give","Sarah", and "Mary". The subject was giventhe fig when she produced the sentence "Marygive fig Sarah", which she did correctly oneight of the first 10 trials, never using the in,correct form for fig but twice using other es-tablished fruit names, perhaps as a request forthese fruits (a methodological problem notedearlier). She was equally proficient when laterexactly the same training procedure was ap-plied to crackerjacks. Notice that the negativetrials-on which she is told that X is not thename of Y-serve to rule out the possibilitythat the name is conferred simply by thegeometry, by the physical contiguity betweenthe language element and the object. Bothwhen it is asserted that X is the name of Y,and when it is denied, the spatio-temporal re-lation between the language element and theobject are identical. Yet only in the case ofassertion does the subject go on to use thelanguage element as the name of the desig-nated referent.

Internal Representation: When Is aPiece of Plastic a Word?When does a piece of plastic cease to be a

piece of plastic and become a word? We mightanswer by saying, when it is used as a word:when it occurs along with other words of ap-propriate grammatical class in sentences, andwhen it occurs as the answer or part of theanswer to questions. For example, we considera small piece of blue plastic to be the namefor apple because (i) it is the word used when,for example, the subject requests apple, and(ii) it is the answer given when the subject isasked, "What is the name of apple?" This is astandard treatment, and we cannot improveupon it, though we may be able to add to it.We might say in addition that the piece ofplastic is a word when the properties ascribedto it are not those of the plastic but are those

of the object designated by the piece of plas-tic. By what means can we determine whetherthis condition obtains?This can be done quite directly by using

match-to-sample procedures to obtain inde-pendent-features analyses of both the wordand its referent. For example, to obtain a fea-tures analysis of apple the subject was given aseries of trials on which an actual apple waspresented along with two alternatives. Hertask was to indicate which of the two alterna-tives was more like the apple than the other.The most explicit version of this test wouldmake use of such words as "similar" or "morelike than" but these words were not a part ofher vocabulary at the time, and we did notfind it necessary to instill them before doingthe present tests. The subject's disposition tomatch-to-sample was strong enough so thatthe three items had only to be set before herin the established geometrical arrangement;this was sufficient to assure that she wouldselect one of the alternatives. The alternativesused in the present tests were: a red plaque vsa green one; a square plaque vs a round one;a square plaque with a stem-like protruber-ance vs a plain square one; and a squareplaque with protruberance vs a plain roundone. The alternatives could be words, ratherthan objects instancing the properties namedby the words. That is, the subject could be re-quired to decide whether the apple was morelike the words "red" and "green" rather thanthe red and green plaques. We did not usewords because of limitations in Sarah's vocab-ulary.

After obtaining a features analysis of apple,we repeated the test exactly except for replac-ing the object apple with the name for apple.Once again the subject was requird to indi-cate whether the sample-now a piece of blueplastic-was, for example, red or green, roundor square, etc. Although the sample was nolonger a shiny red apple but a piece of blueplastic, the subject assigned to the plastic thesame properties she earlier assigned to theapple (see Table 1). Surely if we did notknow that the plastic stood for apple, wewould be confused by her analysis of it; wemight reasonably conclude that she did notunderstand matching-to-sample. But this isruled out by her analysis of the object apple,which accords nicely with the human analysis.The properties she assigns to the word are

123

DAVID PREMACK

immediately sensible if we consider that heranalysis of the word is not of its physical formbut of that which the form represents.

Bertrand Russell in an early work (1940)noted that uttering a word is like jumping inthat both the word and the jump are responses,

but that the word differs from the jump inthat it has meaning. Skinner (1957) com-

mended the first part of Russell's observation,but rejected the second part on the groundsthat a behavioral analysis of language shouldeschew or at least does not need the conceptof meaning. Since I can think of no philoso-pher more admirable than Russell, nor any

psychologist whose reputation is more envi-able than Skinner's, I am pleased to suggestthat they are both wrong: Skinner in holdingthat we do not need meaning, and Russell in

suggesting that the distinction between re-

sponses with and without meaning is a radicalone. On the other hand, there is a sense inwhich both men are right, and their correctjudgments lead more surely to the points thatneed attention now than their more falliblejudgments.

Russell is correct in arguing that a responsemay represent something other than itself. Theidentical-features analysis which the chimp as-

cribes to apple on the one hand and to thename of apple on the other supports that con-

clusion directly. But Skinner is correct in urg-ing that the process by which a response be-comes a word is not a unique one, not a proc-

ess different in principle from the one bywhich a pigeon learns to peck a key when itis lighted. Words do not require special train-

T A B L E I

FEATURES ANALYSES OF APPLE

OB ECT

I

AND..

APPLE

WORD

iLL ___+ 4'~~~~~~~~~

EM I . i. I

124

A FUNCTIONAL ANALYSIS OF LANGUAGE 125

ing metlhods. The procedures that train ani-mals will also produce words.

REFERENCESChomsky, N. Aspects of the theory of syntax. Cam-

bridge: MIT Press, 1965.Hewes, G. W. Lateral dominance, culture, and writ-

ing systems. Human Biology, 1949, 21,. 233-245.

Lenneberg, E. Biological foundations of language.New York: Wiley, 1967.

Premack, D. and Schwartz, A. Preparations for dis-cussing behaviorism with chimpanzee. In F. Smithand G. A. Miller (Eds.), Genesis of language. Cam-bridge: MIT Press, 1966. Pp. 295-335.

Russell, B. An inquiry into meaning and truth. Lon-don: George Allen, 1940.

Skinner, B. F. Verbal behavior. New York: Appleton-Century-Crofts, 1957.