- Email: [email protected]
Words: What are they, and do animals have them?
Cognition, 37 (1990) 197-212
University of Pennsylvania arch 18, 1988, final revision accepted June 12, 1990
Abstract Premack, D., 1990. Words: What are they, and do animals have them? Copition,
37: 197-212.
Since the word is not a well-defined entity like the sentence, one looks for findings that may help to clarify it. The effect of nonsense words on the young child’s sorting of taxonomic versus thematic alternatives is said to be such a finding. A young child given, say, duck as a sample, goose and nest as alternatives, picks nest (thematic alternative), whereas the older child picks goose (taxonomic). However, if told the duck is called “ZLT” in Croatian, and asked to “find another ZLT”, the young child shifts to goose. Markman and Hutchinson (1984) claim this demonstrates that young children know that words are ‘names oj’ object categories” (and that this knowledge protects them against false hypotheses, facilitating their acquisition of words). In the present study, we applied the Markman et al. procedure to young “language-trained” chimpanzees. The animals were at an early stage of training, having used “words ” solely in the function “X goes with Y”, or “if shown X, get I”‘. lthough these func ns are notably weaker than ‘*X is the name of a cdztecc-t~onomic shift, thus behaving like gory ‘)9 the animal’s showed a thermat:young children. The Markman-Hutchinson interpretation of the shift effect is unsatisfactory in two respects: (1) the shift effect can be explained without attributing any knowledge of what a word is 5~ either creature, child or ape; more important (2), the interpretation does not address the main question: what is a “name” and what does a child think it is? We conclude with a discussion of what a word is, appealing to information retrieval on the one hand, and intention to refer on the other.
*Supported by a grant from the McDonnell Foundation. My thanks also to EcoIe Polytechnique, CREA, Paris, with which I was associated during the period in which i rewrote this paper. Finally, it is a pieasure to acknowledge my indebtedness to Ann James Premack with whom I have discussed the present ideas and many others over the course of a long period. Reprint requests should be sent to David Premack, Psychology Department, University of Pennsylvania, Philadelphia, PA 19174, U.S.A.
OOlO-0277/90/$5.30 0 1990- Elsevier Science Publishers B .V.
Do animals have them‘!
t questions. The former is . If we could answer it, we d, however, we have no answer 11-definedentity; in contrast to e formal apparatus of linguise word often turns up in han most foundlings. For as a conditioned response word turns up as an If a primitive mental e word continues to look for lodgings where it can count on clear and consistent treatment. Quine’s remark of 30 years ago “. . . what counts as a word . . . is less clear than what counts as a sentence” (1960) has not lost its relevance today. The complexity of words ;:bn”r be reduced by distinguishing between first or early words - words as seni-isolated items - and words as the constituents of sentences. As building blocks, words become interconnected, forming a semantic structure, one that maps an underlying conceptual structure. But in the early speech er sign of children (and perhaps in that of chimpanzees as well), words do not yet have this character. This is not to say that the “thousands of concepts” (Carey, 1985; Keil, 1983) with which one credits the child (and perhaps the chimpanzee) are not already in place; mental life does not begin with language. But the puzzling manner in which all languages contact the sume conceptual structure (Slobin, 1985; Talmy. 1946, 1983), and ye; cut the structure into different categories owerman, 1988 for recent examples of this linguistic nonequivalence), s yet to appear. It is a simpler period. Words “name” certain things - for example, objects; actions (but not yet mental states), properties, persons. Two questions are outstanding in this simpler period. What is it that words n ? And even more, what is meant by “‘name”? nd her colleagues have already offered an answer to the first categories of similar objects” (Markman & Hutchp- 23). Children know this, eve very young ones, and this knowledge protects them against false hypotheses, enabling them to learn words efficiently. For instance, when a mother says “dog” while pointing to a dog, the child rejects brown, barking, furry, etc. as possible referents, acceptiq only dog as an exemplar of a category. Markman and Hutchinson (1984) base this claim on the interesting effect that novel words have on the thematic-taxonomic shift: Young children tend to group objects according to thematic relations, e.g., they might put a dog and a bone together because the dog eats the bone. In contrast, I;_’s!mm a dog and told it is called “a dax,” and then told to “find another dax,”
children now tend to put a dog and a cat together . . . Thus, the presence of a novel word causes children to seek taxonomic rather than thematic relations. (PO1221
Now, of course, some words do indeed name (whatever that is) object categories, and children may even know that. But we may still raise this question: is the effect of novel words on the thematic-taxonomic shift actually evidence for that claim? Could a shift effect be shown for reasons quite different from - in particular, weaker than - those Markman et al. propose? One way to deal with that question is to apply the Markman et al. test to nonhuman species. Especially if nonhumans too show the shift effect, we may wish to consider weaker explanations of the effect. Moreover, the fact that, when learning words, children apparently give objects a higher priority than other alternatives - properties, parts of objects, actions or the like - can be explained simply by granting the child a hierarchy of perceptual preferences. Such an assumption does not require crediting the child -with any knowledge about words at all; furthermore, though the assumption is obviously post hoc, it can be tested independent of words. But the Markman et al. discussion of the shift effect is dissatisfying for a more basic reason. In claiming that children know that words name object categories, we are not told what “name” is. Thus, our main question - what is a word? - is left unanswered. What is the nature of the “name of” relation, how does it differ from other relations that may hold between an item and a category, and can a child discriminate between the several possible relations? When a child knows that “a word names a category”, what is it that he knows? We will return to these questions in the discussion section. Is the thematic-taxonomic shift species specific? In the present study, we applied the Markman et al. procedure to four chimpanzees. That is, we gave them match-to-sample tests, in some cases accompanying the sample with a novel or nonsense word and in other cases not. Chimpanzees could be tested in the same manner as children for two reasons. impanzees too show a developmental change in the sorting of objects, sorting largely on a thematic basis when young, on a taxonomic basis when older (why either species shows this change is an unanswered question).’
‘Smiley and Brown (1979) in an instructive study, found that, whereas young children sorted thematically and older ones taxonomically, adults of 60 and older reverted to the thematic approach! This both shows that language competence is not a sufficient condition for taxonomic sorting, and puts Sarah’s results in an interesting perspective. Is language a necessary condition for taxonomic sorting.3 Although 13 years of disuse had virtually eliminated the “language” Sarah had been taught, she uonetheless sorted taxonomically. Taxonomic sorting may depend less on language than on a preference to compute equivalences on a relational rather than absolute basis. Now language training promotes computation of relations between relations in chimpanzee (Premack, 1988b) - but the computation continues after disuse of language. Since this same computation figures in analogies, one wonders about the spontaneous construction of analogies in adults of 60 and over.
wrnack
Second, chimpanzees, unlike many other nonhuman species, readily acquire . -sample. Ill istinctive histories of the four a imals lent a further benefit to the r animals were both subadult and trained in “words”; study. Two of the to normal young children. If the functional properties e animals’ plastic words resembled those of children, then the addition el words to the target object should have the same effect it has in a shift in sorting from thematic to taxonomic. A third animal was subaduh but not trained in language. As a young animal, it too should sort on a thematic basis, but, having no training in language the addition of novel words to the target object should not bring about a thematic-taxonomic The animal represents an interesting control for the Markman-Hutch hypothesis, one that is not possible with children. Although for a normal 3-year-old child lzot to know words is a contradiction in terms it Ssa perfectly normal condition for the 3-year-old ape. A fourth animal represented another kind of control. Trained in language when young but now adult, it should respond as would a human adult: sort on a taxonomic basis and thus be unaffected by the addition of novel words. Only one control was lacking an adult not trained in language - ordinarily the most common animal of all, but unfortunately one we did not have in the laboratory.2
-“Ihe ~xocedure we used in training these animals was simpler than any we had used before. As the trainer and animal sat on the floor in the middle of a small room, a different object was placed in each of the four corners of the room. The trainer then stuck a plastic word on the writing board, called it to the animal’s attention, and released the animal into the room. The animal’s task was to find the object that “went with” the word and return it to the trainer. When rhe animal chose correctly, it was patted and praised; its incorrect choices were corrected. At the end of the session the animal was fed a small portion of a favorite food. Ultimately, four animals were trained in this way to a 75% correct criterion (chance level = 25%), whereupon we turned to a different question. Did the animal understand the role played by the plastic word in the sense that the trainer did? To answer this question we interpolated trials in which the plastic word, though placed on the writing board, -wasno longer visible to the animal. The board was either turned away from the animal, or rotated completely so that it faced into the trainer’s chest. The animal was in a position analogous to that of a child: while normally the child is told in a clear voice which object to take, subsequently the voice becomes either dim or itiaudible. Under such conditions will a child make a choice despite the lack of information or will she request her informant to “speak up” ? One of the four animals acted despite the lack of information; setting out into the room without hesitation she selected one of the objects and returned with it as she did on any normal trial (naturally performing at cance letel). A second animal whimpered and cried, and though reluctant and complaining ah the while, made its choice nonetheless. The other two, however, went straight to the source of the problem and immediately rectified the loss of information. When the board was merely turned, they twisted their bodies to achieve a view of the word, but when the board was inverted, they took actual hold of it and turned it around (Premack, 1986). Interestingly, in accuracy of association, the least gifted animal - that went about her business uninformed and uncomplaining - was tied for first place (88%’ correct).
%brds
201
s Subjects
Four chimpanzees (Pan troglodytes) served in the study. Three were domestic-born juveniles, 5$ to 6 year;, old, art of a social group that entered the laboratory as infants. The fourth was Sarah, an African-born female, maintained in the laboratory from infancy, and taught an artificial language when a subadult; she was 25 years old at the time of the study. All subjects had extensive training histories, having been taught, for example, match-to-sample as infants. The animals were fed three meals a day and given continuous access to water, and thus were no more food-deprived than middle-class children. Procedure
A standard match-to-sample procedure was used, a sample and two alternatives being presented on each trial, all consisting of ordinary objects with which each subje ct was highly familiar. The animal’s task was to select one of the two alternatives and place it with the sample. No differential feedback was given. Instead, the animal was praised for all its choices and given a small portion of a favorite food at the end of each test session. Six test sessions were run, each consisting of 24 trials. Sessions were given either daily or at 4%hour intervals. Two lists of 12 trials were constructed (see Table 1). Each trial consisted of a sample or target object, and two alternatives - one the taxonomic alternative, the other the thematic. The taxonomic alternative was always a token of the same type as the sample, different only in physical detail. For instance, the sample was a teaspoon, the taxonomic alternative a tablespoon; the sampiea large red apple, the taxonomic alternative a small green one; etc. The thematic alternative was chosen so as to exploit the known history of the animals. For instance, when shoe was the target, shoelace was the thematic alternative, shoes with laces being a common toy; likewise when lime was the target, knife was the thematic alternative since cutting pieces of fruit with a knife was one of the animals’ favorite play activities. The two lists (shown in Table 1) differed in only one respect. In list 1, all the target objects were ones for which the experimental animals (Walnut and Qpal) had been taught names; in list 2, they had not been taughtnames for any of the target objects. This difference had no bearing on either of the control animals, for Frieda had never been taught any names, and Sarah had largely forgotten those she was taught; at the time of the test, it was 13 years since she used any plastic words except for “same/different”.
Table 1.
Lists of test item
_________
.--
--
_-___I_-
Target object A~c~~~ives _---Knife Apple Pen Shoe CUP Carrot ?+dlet spoon Spray bottle DON Sponge scissors
Pen!paper Shoe&ace Cupljuice Carrot/k&e Mallet&rape Spmnkontainer
Sponge/waxer Scissors/paper
2. Clnnamed target objects
Target object Alternatives Paper Lid Coat Sock Nandywiper Lock &tray Jell0 Lime Dough Hose psrt Bubbles --
Hand
Lime/knife D 3-I Bu
ons were run, all of them with list I. The control rget object and the two alternatives without any ect. This served to establish how the individual sorted, r a taxonomic one. In a second test condition, anied by its name, that is, the name taught the stic word being placed alongside the target each target object was accompanied by a nt one for each target. A novel or nonsense he cIass of items that constituted words in the he class consisted of pieces of plastic, varying
Words
Table 2.
203
Test conditions and order in which they were given 1. Control. Sample: Target object alone 2. Sample: Target object + name of target object 3. Sample: Target object (named) + Nonsense word 4. Sample: Target object (unnamed) + Nonsense word 5. Control. Sample: Target object alone 6. Sample: Target object + “foreign” object
in shape, size, and color, each backed with metal so that it adhered to the magnetized language board. These items are nonsense or potential words in the sense that, though each conformed to the word class, they had never been used as a word. In the fourth test condition, each target object was accompanied by a “foreign” object, a different one with each target object. Foreign objects - for example, bottle caps, dominoes, bolts - did not cc>il’iirtilto the properties of the word class and thus could not be considered novel or nonsense words. List 1 was used in each of the four test conditions, and list 2 was used once in a special test condition (see Table 2 and results). The 12 tri:Js pf both list 1 and list 2 were presented twice in every test, the left-right position of the two alternatives being interchanged semi-randomly, giving a total of 24 trials per test. The order in which the test conditions was run is shown in Table 2; the control condition was rtrin tw~e, in the first and fifth position, to establish that the base condition was stable during the test period.
s
Figure 1 shows the number of taxonomic choices each animal made in each of the six tests. Consider first the two control tests: they show that all three young animals - Frieda, Walnut ard Opal - sorted on a thematic basis; whereas Sarah, the one adult an rn~l, sorted on a taxonomic basis. Sarah chose the taxonomic alternative 37 out of 48 trials; the three young animals chose it only an average of 9 out of 48 trials @ > .Ol). Language training did not appear to figure in this difference. Frieda, the nonlanguage-trained control, showed the same degree of thematic sorting as her two language-trained peers. Test 3, in which the target object was accompanied by its name, produced the expected effect. For the two experimental animals, Walnut and Opal, the plastic word that appeared alongside the target object was the name of the target object. Both of these animals came under the control of the plastic
mack
Figure 1.
0
I 2 8 1 CONTROL
3 I
4
5
6 lwkml
words: they shifted to the raxonomic alternative, choosing the object named untrained in words, was naturally unafto choose the thematchange, continuing to trend continued: Frieda (the young animal) chose the n all test conditions, eas Sarah (the older one) chose the taxonomic alternative on all test tions, the choiczs of the two animals being dete ined apparently by their respective ages. tests of greatest interest are the remaining three: 4, 5 and 6. The in t focuses on the two experimental animals, Walnut and Opal. Novel words affected these two young guage-trained chimpanzees in the same way as they do young children. both tests 4 and 5, when the novel or onsense words accompanied the target objects, both Walnut and Opal made redominantly taxonomic choices, an average of 20 out of 24 trials (p > .05). fact, by comparing tests 4 and 5 with test 3, we see that the experimental animals made as many taxonomic choices when given novel or nonsense words as when given the actual names of the target objects. Comparison of tests 4 and 5 adds the interesting fact that the effect of the novel word is felt whether the target object is named or not (target objects were named in test
Words
205
4, unnamed in 5). The marked effect of novel words on the experimental animals contrasts with the complete absence of effect on Sarah a Novel words had no effect on these control animals. Finally, test 6 confirms the legitimacy of speaking of the effect of specifically novel or nonsense words. Foreign objects, items that do not conform to the word class, do not produce a taxonomic shift; under this condition, both Walnut and Opal returned to the thematic alternative. This control, which incidentally is lacking in the child research, demonstrates that it is not simply the addition of any object (o*1 Dyuaen ment sound) that produces the shift, but only objects with wordlike properties. In brief? only animals that were both subadult and language-trained showed a shift effect, one that was as pronounced for novei words as for actual words, and one that did not occur at all for foreign objects.
The present results indicate that novel words lead to a thematic-taxonomic shift in the chimpanzee as they do in the child; the effect is not species specific. One can always maintain, of course, that, while children; and chimpanzees act the same, they do so for different reasons. This does not seem t& ideal occasion for such a claim, however, since there is no apparent reason for making it. On the other hand, the fact that the two species share the shift effect does 3ot wowu.tiauir #%?*Qi%l;C~ any deep correspondence beta-&en human and animal words. As suggested earlier in the Introduction, the shift effect does not tell us what a word is - even what a child (or ape) thinks it is - and there is reason to suspect that tlb * effect is a superficial one. For example, the shift effect is not only found in animals but in animals at an early stage of language training. Walnut and Opal had been taught 25 plastic words, primarily names of objects. The names of ~WG actions had only recently been introduced, and training in object-action pairs had just begun. Moreover, even this may be more than is needed; animals given may still show the shift effect. What is the minimal knowledge an individual must have to show the effect? We may find an answer to this question by noting that the effect of placing a novel word alongside an object is exactly like that of placing thti actual name of the object there. Both lea the individual to take the taxonomic alternative. For instarce, when Opal and Walnut were given a large knife as the target, a small knife and an apple as alternatives, both started by choosing the apple. Adding the plastic word “knife” - placing it alongside the target
- shifted both anim
;
they mw to
se, “knife’” is after
all the name of knife - yet this out@ than thematic associations; they exe titian with thematic associations, words response. ct; when placed alongside the target Novel words have exactly the ssociation, no less effectively than object, they too overcome the rds that are placed actual words. Thus, the animal appears to treat ct, no less so than alongside the tar bject as being the n y? Some previous data suggest an answer to this questhe actual name. tion. Sarah was taught many words with the instruction “X name of Y:t, X being a potential word (an unused piece of plastic), Y an unnamed object. But with the continual need for new words - every transfer test required several, an tests of this kind re given daily - the trainer found “X name of Y” tedious. She devised a si ler procedure. She took advantage of two containers of objects, one a 1 km of unnamed objects that were reserved as referents for new words, the other a supdv of appropriate pieces of plastic cut from long molds of plastic and back&i with a piece of metal so that the word would adhere to the magnetized writing board. Taking one of the not-yet named a3bjects in one hand, and one of the not-yet used words in the other, she held bath of them up f-VI the animal’s inspection. She then clapped the two tohenceforth found that words introduced in this manner could be name of thz object no less effectively than words introduced in a onal manner. a novel word and object together, on the one hand, and placing alongside the target object, on the other, appear to have the he animal treats the novel word as the “name of” the object in question. A naive chimpanzee would not do so, of course - only one that has been appropriately trained. n fact, the naive chimpanzee has great difficulty learning the first words ht it. Three hundred or more trials are often needed to introduce the animal to words. owever, once the animal has acquired first words, it acquires subsequent words in fewer trials, ultimately in a single trial. What does the animal learn that accounts for this transition? A series of tests showed that, in the course of learning its first words, the animal had learned to recognize members of the two classes - those of the word, on the one hand, and those of the referent, on the other. The animal could distinguish novel or unfamiliar members from nonmembers in both cases (all referents in this case were fruit and thus had class properties). Besides a knowledge of class properties, the animal had also learned the appropriate use of each class; for
example, to use plastic words to obtain Btot vice versa; to place words on the writing board, not in the m (Premack, 1976). At this early stage of training, astic words were virtually confined to a single use: to “request” fruit from
e trainer. More interesting language-like uses of the plastic words had yet to occur. For instance, plastic words were t yet used to answer questions, for example, “? color of apple” (What is color of apple?), “square ? cracker” (What is the relation between square cracker?), etc., which the animal “answered” by removin interrogative particle and replacing it with the appropriate word. But s the point - “advanced” language-like uses of the plastic words are not needed in order to produce either a shift effect, or one-trial learning. Both of these effects can be produced merely.by arranging for the individual to experience a co+ sistent functional, relation between the individual members of two different classes. If a member of one class can be used to obtain a member of the other class, and this holds for all members of both classes, the generalization leading from this experience will produce both one-trial learning and a shift effect. That is, this experience produced these effects iq11appropriate creatures, apes (and presumably children) being evidently appropriate. X can be used “to obtain” Y, though not equivalent to X is the “name of” Y, nonetheless produced a shift effect; we cannot therefore take the shift effect as evidence for the knowledge that a word is the “name of” a category. Note we are not arguing that the child’s experience was as narrow as that of the apes - restricted to the use of words as requests; certainly it was not. But children do use words to make requests, and since this narrow use was scfficient to produce a shift effect in the ape, we have no reason to suppose that it is not sufficient to produce the same effect in the child. This is the real problem: for all we know, the experience “X can be used to obtain Y” may be one of a small family of experiences that leads a child (or even an ape!) to treat “X and Y” as exemplars of the primitive relation “X name of Y”. But, of course, we cannot simply assume that - it is the most essential part of the problem we are trying to solve. Unfonrlnately, as should now be clear, the shift effect provides no leverage in solving that problem. Note that the effect of the novel word held 8s -:dell for target objects that were already named as it did for unnamed target objects (compare results of tests 5 and 6), casting doubt on another claim of Markman and Wachtel (1988), namely, that of “mutual exclusivity”. According to these writers, children acquire -words efficiently not only because of the assumption that words name categories (“taxonomic organization”), but also because of the assumption that if an object has o e name it will not have another (“mutual exclusivity”). However, if children hold this assumption, they hold it lightly,
for the assumption does
shift effect. For example, when d this is a “dax, find me another dax”, the shift effect induced is not confined to unn ed objects. Although “dax” may be the second: name the c ild is given for the object, it still produces the shift ees further resemble one another in that effect. Thus children an the use of unnamed objects. the shift effect does not hift effect really depends on the assumpe specifically eat@gories. ust the target object be an ti exemplar of a category? To be sure, in all tk amples at ha&. in the ape no less than in the child, target objects are of exactly this kind. Dolls, whistles, apples, knives, etc. are members of categories. However, that is exactly the problem: we lack tests in which the target object is not the member of a order to be certain that category membership is critical, tests are hich the target objects are not traditional members of categories, and novel words therefore do not name categories. Moreover, we must find that in such tests novel words do not lead to a shift effect. Since there no actual tests of this kind, we may fall back on a Gedankenexperimen ur Gedankenexperiment is exactly like all other shift effect tests except for one feature: the target object is not a category member. Instead, we use a picture of the child’s mother as the target object. As alternatives, we offer the standard taxonomic and thematic choices; for example, another somewhat different picture of the mother, and a picture of a familiar purse that the child easily recognizes as belonging to the mother. Let us se that the child shows thematic sorting in this case too, choosing the er’s purse rather than her picture. This assumption will enable us ito take e next step, and inquire about the effect of novel words. odeling our instructions as much as possible on those used by Markman we tell the child “Nina is what your mother was called when she was and ask her to “find another Nina”. Does this produce a taxonomic shift, leading the child to shift from picture of purse to that of f it does not, then all is well for the Markman et al. hypothesis: a” is not the name of a category, and therefore it should not produce a shift effect. On the other hand, if “Nina” has the usual effect of the novel word, then all is not well for the hypothesis. The child grasps, I assume, that “Nina”, as the proper (childhood) name of the mother, is not the name of a category (though probably the child’s understanding of this distinction merits study in its own right). In any case, in all tests of the shift effect only traditional category objects have been used as target objects, creating the imgression that the effect depends upon the knowledge that “words naive categories”. This is, I think, an unlikely assunlption and, in any case, an untested one. not
interfere
tit
the
words
209
What is a word? Information retrieval and reference Even a patient reader may now complain, “This article is ending without having made any progress on the principal question: What is a word? What is meant by ‘name of?” Let us consider two properties of words, one pointed out by behaviorism, the other by mentalism. First behaviorism: words are information-retrieval devices. Traditional discussions of language make the same point when they treat displacement as a hallmark of language. The ability to talk about things that are not present depends, of course, on the information-retrieval capacity of words. What does information-retrieval depend on? Mainly on a certain quality of mental representation. A species that suffered a major disparity between what it can perceive and what it can represent - a species in which representation did not do justice to its perception - could not talk effectively about what it cannot see. Words have no power as such; they borrow their power from the representations they retrieve. Were the chimpanzees’ plastic words effective information-retrieval devices? Dividing this question into two parts, l:re first established the nature of the chimpanzee’s mental representation, and then tested its ability to retrieve these representations with the plastic words. Match-to-sample tests were used in both cases. For example, the animal was given a taste of, say, peach, and then shown a green patch and a yellow one (or a paper silhouette of an apple and a peach, a peach pit and an apple seed, etc.). Three of the animals chose the correct alternative over 80% of the time; Sarah performed still better, successfully matching each of seven different features into which eight fruits were divided with every other feature (Premack, 1976,. The main point was established: there was no major disparity between perception and representation in the chimpanzee; representation did justice to perception. We then tested the ability of the plastic words to retrieve the information we knew the animals had stored. Plastic words were substituted for the former alternatives in the match-tosample tests. For instance, the animal was shown an apple stem as the sample., and offered as alternatives the plastic words “apple” and “peach”. With words as alternatives, three of the animals passed tests that they had failed earlier, when physical features were used as alternatives. In effect, the information retrieved by the plastic words exceeded that retrieved by actual parts or features of the fruit. This property of the plastic words will account for the animal’s ability to acquire words noncstensively. Sarah was taught the word “brown” with the instruction “brown color of chocolate”, the words “color of’ and “chocolate” being already established. Neither chocolate nor any brown object was pres-
ent. Next, she was given the instruction “take brown”, in the presence of four discs only one of which was brown, and chose correctly (Premack, 1976). ne can say of the information-retrieval side ef words that it is of little t does not seem quite for by conditioning. interest since it is we have already seen one accounts for it the right reaction. that a certain quality of mental representation is a major prerequisite - the sheer amount of information controlled by the Kxd is uncanny. Perhaps one can put this side of wor into perspective by considering what happens when opriately train chimpanzee is given the instruction: “X name of Y”. en so instructed, the animal’s ment presentation of the object Y is can be evoked mentally by X. red to the new word X. Hencefo s seems a remarkable transformation, an uncanny characteristic of mind. regrets to the readers for whom this property is demeaned by the fact that it can be accounted for by conditioning. mentalism we receive a second property of words, namely words ot all words but at least those that name objects. Here we could amara (1982), who treats reference as a special intentional state, one which young children can recognize. They can distinguish, he claims, when a speaker says “Betty” or “dog”. ~tc. w intention to refer to the item in question, and when, in contrast, the speak s simply musing aloud. Then our question becomes: is this special intentional state one that is recognized and/or instantiated by any nonhuman species? One way or another we need so we need to present both children and apes with speakers who employ the same words, in some cases with intent to refer, and in other cases without, and then determine whether either or both species can make the proper distinctions. Children, it goes without saying, will be far better at this discrimination than apes; indeed, there must be circumstances children understand nicely that apes cannot grasp at all. But are there no circumstances under which the ape can make this discrimination? That does not go without rtunately, there appears to be no direct evidence. The closest evidence concerns whether chimpanzees attribute any states of mind to others (Premack, 1988a; Premack & Dasser, in press; Premack & Woodruff, 1978). On the whole the evidence is positive, though there is ample indication that the animal’s competence is limited, that its “theory of mind” is not equivalent to the human one (Premack, 1988a). Nevertheless, if an individual attributes SOme kinds of intentions to others, as the chimpanzee apparently does, might it not then attribute specifically intention to refer? There is a chance. How similar is intentional reference to other kinds of intentions? We are, I think, beyond the polemical stage of this discussion; we see what some of the ques-
tions are and know how to test them (even as we know how to test for “theory of mind” in general). Children use words in ways that apes do not, one or more of which may concern reference. One that is especially suspicious is the strange reaction that children between approximately 10 and 13 months of age show to an object that is of special interest to them. For a period of perhaps two weeks the sight of the object produces an exceptional arousal state in the child. The arousing object in the case of a l(\.S-month-old male was a moving truck, a moving bus in the case of a 13.month-old male, and a swimming goldfish in the case of a 12-month-old fe ale. Each child pointed excitedly at the moving object, calling out the name repeatedly - %l”, 66bu!9S,“ef !” - all the while avidly seeking eye contact with its recipient. Sometimes the child had a favorite recipient, and endangered its life in an attempt to communicate with her, as when one child (riding in the front with its father) sought to climb out the window of a truck to reach its mother riding in the back. What was the meaning of the child’s behavior? Was it a request? Did the child want to reduce the distance between itself and the arousing object? Did it want to touch or at least see the object up close? One could test this more practically with the fish than with either the truck or bus. One took the fish bowl down from the shelf and offered it to the child. The effect was clear from the first test: the child stepped back, rejecting the opportunity to reach in and touch the fish, or even to draw near and observe it more closely. Her pointing arm dropped and she fell silent. Moreover, once the bowl was restored to its proper place the child immediately reverted to its former behavior - shouting (“ef! ef!“), pointing, and seeking eye contact. The child’s message was not that of a request for the object. This behavior had several other properties in all three children (although at least five children are needed for statistical significan , family size is rarely designed with statistics in mind). Each child behav in this manner with respect to only one object, for only a short period (perhaps 10 days), using a word which, though an early word, was not the child’s first word. For example, all three children used “mamma” at this stage. But no child behaved in this way with respect to the would-be referent here, that is, pointed at mother, shouted “mama!” while avidly seeking eye contact with father or some other recipient. Was “mama” merely a pleasing sound the child emitted in the presence of the appropriate party, while ‘Yu”, “bu” and “ef” were the real first words, that is, words that referred‘? This is a tantalizing question but not a good one because there is no apparent way to answer it. Finally, we see there is no incompatibility between the properties that behaviorism and mentalism emphasize. Any words an individual uses in acts of reference will, of course, be well associated with her mental representation
21
remtack
e objects to which she refers. trieval.
ce presupposes information re-
Bowerman, M. (1988). Inducing the latent structure of language. In F.S. Kessel (Ed.), The development of language and language reseurchers: Essays in honor ofRoger Brown. Hillsdale, NJ: Lawrence Erlbaum. Carey, S. (1985). Concepml change ila childhood. Cambridge, MA: MIT Press. Keil, F.C. (1983). Semantic influences and the acquisition of word meaning. Berlin: Springer-Verlag. Macnamara, J. (1982). Names for things. Cambridge, MA: MIT Press. Markman, E.M., & Hutchinson, J.E. (19s). Children’s sensitivity to constraints on word meaning: Taxonomic vs thematic relations. Cogrritive Pq&oiogy, 16, i-27. Markman, E.M., & Wachtel, G.F. (1988). Children’s use of mutual exclusivity to constrain the meanings of words. Cog&ive Ps@roiogy, 20, 121-157. Premack, 1). (1976). intelligence in ape and mun. Hillsdale, NJ: Lawrence Erlbaum. Premack. D. (1986). Guvagai? or the future history of the animal ianguage controversy. Cambridge, !:A: MIT Press. Premack. D. (198Sa). “Does the chimpanzee have a theory of mind?” revisited. In R.W. Byrne & A. Whiten (Eds.), Machiavellian intelligence. Oxford: Oxford University Press. Premack, D. (198Sb). Minds with and without language. In L. Weiskrantz (Ed.), Thought without language. Cambridge, UK: Cambridge University Press. Premack, D., & Dasser, V. (in press). Perceptual origins and conceptual evidence for theory of mind in apes and children. In A. Whiten (Ed.), Mindrecxding: The evolution, development and simulation of secondorder mental representations. Oxford: Basil Blackwell. Premack, D., & Woodruff, 6. (1978). Does the chimpanzee have a theory of mind? The Behavioral and Brain Sciences, I, 515-526.
Quine, W.V. (1960). Word and object. Cambridge: MIT Press. Skinner, B.F. (1957). iJJerbrrlbchbvior. New York: Appleton-Century-Crofts. Slobin, D.1. (1985). Crosslinguistic evidence for the language-making capacity. In D.I. Siobin (Ed.), The crosslinguistic study of language acquisition. Hillsdale, NJ: Lawrence Erlbaum. Smiley, S.S., 8 Brown, A.L. (1979). Conceptual preference for thematic or taxonomic relations: A nonmonotonic age trend from pre-school to aid age. Journal of Experimental Child Psychology, 28, 2@257.
Talmy, L. (1976). Semantic causative types. In M. Shibatami (Ed.), Syntax and semantics, Vol. 6: The grammar of causative constructions. New York: Academic Press. Talmy, L. (1983). How language structures space. In H.L. Pick Jr. & L.P. Accredobo (Eds.), Spatial orientution: Theory, research and application. New York: Plenum Press.
University of Pennsylvania arch 18, 1988, final revision accepted June 12, 1990
Abstract Premack, D., 1990. Words: What are they, and do animals have them? Copition,
37: 197-212.
Since the word is not a well-defined entity like the sentence, one looks for findings that may help to clarify it. The effect of nonsense words on the young child’s sorting of taxonomic versus thematic alternatives is said to be such a finding. A young child given, say, duck as a sample, goose and nest as alternatives, picks nest (thematic alternative), whereas the older child picks goose (taxonomic). However, if told the duck is called “ZLT” in Croatian, and asked to “find another ZLT”, the young child shifts to goose. Markman and Hutchinson (1984) claim this demonstrates that young children know that words are ‘names oj’ object categories” (and that this knowledge protects them against false hypotheses, facilitating their acquisition of words). In the present study, we applied the Markman et al. procedure to young “language-trained” chimpanzees. The animals were at an early stage of training, having used “words ” solely in the function “X goes with Y”, or “if shown X, get I”‘. lthough these func ns are notably weaker than ‘*X is the name of a cdztecc-t~onomic shift, thus behaving like gory ‘)9 the animal’s showed a thermat:young children. The Markman-Hutchinson interpretation of the shift effect is unsatisfactory in two respects: (1) the shift effect can be explained without attributing any knowledge of what a word is 5~ either creature, child or ape; more important (2), the interpretation does not address the main question: what is a “name” and what does a child think it is? We conclude with a discussion of what a word is, appealing to information retrieval on the one hand, and intention to refer on the other.
*Supported by a grant from the McDonnell Foundation. My thanks also to EcoIe Polytechnique, CREA, Paris, with which I was associated during the period in which i rewrote this paper. Finally, it is a pieasure to acknowledge my indebtedness to Ann James Premack with whom I have discussed the present ideas and many others over the course of a long period. Reprint requests should be sent to David Premack, Psychology Department, University of Pennsylvania, Philadelphia, PA 19174, U.S.A.
OOlO-0277/90/$5.30 0 1990- Elsevier Science Publishers B .V.
Do animals have them‘!
t questions. The former is . If we could answer it, we d, however, we have no answer 11-definedentity; in contrast to e formal apparatus of linguise word often turns up in han most foundlings. For as a conditioned response word turns up as an If a primitive mental e word continues to look for lodgings where it can count on clear and consistent treatment. Quine’s remark of 30 years ago “. . . what counts as a word . . . is less clear than what counts as a sentence” (1960) has not lost its relevance today. The complexity of words ;:bn”r be reduced by distinguishing between first or early words - words as seni-isolated items - and words as the constituents of sentences. As building blocks, words become interconnected, forming a semantic structure, one that maps an underlying conceptual structure. But in the early speech er sign of children (and perhaps in that of chimpanzees as well), words do not yet have this character. This is not to say that the “thousands of concepts” (Carey, 1985; Keil, 1983) with which one credits the child (and perhaps the chimpanzee) are not already in place; mental life does not begin with language. But the puzzling manner in which all languages contact the sume conceptual structure (Slobin, 1985; Talmy. 1946, 1983), and ye; cut the structure into different categories owerman, 1988 for recent examples of this linguistic nonequivalence), s yet to appear. It is a simpler period. Words “name” certain things - for example, objects; actions (but not yet mental states), properties, persons. Two questions are outstanding in this simpler period. What is it that words n ? And even more, what is meant by “‘name”? nd her colleagues have already offered an answer to the first categories of similar objects” (Markman & Hutchp- 23). Children know this, eve very young ones, and this knowledge protects them against false hypotheses, enabling them to learn words efficiently. For instance, when a mother says “dog” while pointing to a dog, the child rejects brown, barking, furry, etc. as possible referents, acceptiq only dog as an exemplar of a category. Markman and Hutchinson (1984) base this claim on the interesting effect that novel words have on the thematic-taxonomic shift: Young children tend to group objects according to thematic relations, e.g., they might put a dog and a bone together because the dog eats the bone. In contrast, I;_’s!mm a dog and told it is called “a dax,” and then told to “find another dax,”
children now tend to put a dog and a cat together . . . Thus, the presence of a novel word causes children to seek taxonomic rather than thematic relations. (PO1221
Now, of course, some words do indeed name (whatever that is) object categories, and children may even know that. But we may still raise this question: is the effect of novel words on the thematic-taxonomic shift actually evidence for that claim? Could a shift effect be shown for reasons quite different from - in particular, weaker than - those Markman et al. propose? One way to deal with that question is to apply the Markman et al. test to nonhuman species. Especially if nonhumans too show the shift effect, we may wish to consider weaker explanations of the effect. Moreover, the fact that, when learning words, children apparently give objects a higher priority than other alternatives - properties, parts of objects, actions or the like - can be explained simply by granting the child a hierarchy of perceptual preferences. Such an assumption does not require crediting the child -with any knowledge about words at all; furthermore, though the assumption is obviously post hoc, it can be tested independent of words. But the Markman et al. discussion of the shift effect is dissatisfying for a more basic reason. In claiming that children know that words name object categories, we are not told what “name” is. Thus, our main question - what is a word? - is left unanswered. What is the nature of the “name of” relation, how does it differ from other relations that may hold between an item and a category, and can a child discriminate between the several possible relations? When a child knows that “a word names a category”, what is it that he knows? We will return to these questions in the discussion section. Is the thematic-taxonomic shift species specific? In the present study, we applied the Markman et al. procedure to four chimpanzees. That is, we gave them match-to-sample tests, in some cases accompanying the sample with a novel or nonsense word and in other cases not. Chimpanzees could be tested in the same manner as children for two reasons. impanzees too show a developmental change in the sorting of objects, sorting largely on a thematic basis when young, on a taxonomic basis when older (why either species shows this change is an unanswered question).’
‘Smiley and Brown (1979) in an instructive study, found that, whereas young children sorted thematically and older ones taxonomically, adults of 60 and older reverted to the thematic approach! This both shows that language competence is not a sufficient condition for taxonomic sorting, and puts Sarah’s results in an interesting perspective. Is language a necessary condition for taxonomic sorting.3 Although 13 years of disuse had virtually eliminated the “language” Sarah had been taught, she uonetheless sorted taxonomically. Taxonomic sorting may depend less on language than on a preference to compute equivalences on a relational rather than absolute basis. Now language training promotes computation of relations between relations in chimpanzee (Premack, 1988b) - but the computation continues after disuse of language. Since this same computation figures in analogies, one wonders about the spontaneous construction of analogies in adults of 60 and over.
wrnack
Second, chimpanzees, unlike many other nonhuman species, readily acquire . -sample. Ill istinctive histories of the four a imals lent a further benefit to the r animals were both subadult and trained in “words”; study. Two of the to normal young children. If the functional properties e animals’ plastic words resembled those of children, then the addition el words to the target object should have the same effect it has in a shift in sorting from thematic to taxonomic. A third animal was subaduh but not trained in language. As a young animal, it too should sort on a thematic basis, but, having no training in language the addition of novel words to the target object should not bring about a thematic-taxonomic The animal represents an interesting control for the Markman-Hutch hypothesis, one that is not possible with children. Although for a normal 3-year-old child lzot to know words is a contradiction in terms it Ssa perfectly normal condition for the 3-year-old ape. A fourth animal represented another kind of control. Trained in language when young but now adult, it should respond as would a human adult: sort on a taxonomic basis and thus be unaffected by the addition of novel words. Only one control was lacking an adult not trained in language - ordinarily the most common animal of all, but unfortunately one we did not have in the laboratory.2
-“Ihe ~xocedure we used in training these animals was simpler than any we had used before. As the trainer and animal sat on the floor in the middle of a small room, a different object was placed in each of the four corners of the room. The trainer then stuck a plastic word on the writing board, called it to the animal’s attention, and released the animal into the room. The animal’s task was to find the object that “went with” the word and return it to the trainer. When rhe animal chose correctly, it was patted and praised; its incorrect choices were corrected. At the end of the session the animal was fed a small portion of a favorite food. Ultimately, four animals were trained in this way to a 75% correct criterion (chance level = 25%), whereupon we turned to a different question. Did the animal understand the role played by the plastic word in the sense that the trainer did? To answer this question we interpolated trials in which the plastic word, though placed on the writing board, -wasno longer visible to the animal. The board was either turned away from the animal, or rotated completely so that it faced into the trainer’s chest. The animal was in a position analogous to that of a child: while normally the child is told in a clear voice which object to take, subsequently the voice becomes either dim or itiaudible. Under such conditions will a child make a choice despite the lack of information or will she request her informant to “speak up” ? One of the four animals acted despite the lack of information; setting out into the room without hesitation she selected one of the objects and returned with it as she did on any normal trial (naturally performing at cance letel). A second animal whimpered and cried, and though reluctant and complaining ah the while, made its choice nonetheless. The other two, however, went straight to the source of the problem and immediately rectified the loss of information. When the board was merely turned, they twisted their bodies to achieve a view of the word, but when the board was inverted, they took actual hold of it and turned it around (Premack, 1986). Interestingly, in accuracy of association, the least gifted animal - that went about her business uninformed and uncomplaining - was tied for first place (88%’ correct).
%brds
201
s Subjects
Four chimpanzees (Pan troglodytes) served in the study. Three were domestic-born juveniles, 5$ to 6 year;, old, art of a social group that entered the laboratory as infants. The fourth was Sarah, an African-born female, maintained in the laboratory from infancy, and taught an artificial language when a subadult; she was 25 years old at the time of the study. All subjects had extensive training histories, having been taught, for example, match-to-sample as infants. The animals were fed three meals a day and given continuous access to water, and thus were no more food-deprived than middle-class children. Procedure
A standard match-to-sample procedure was used, a sample and two alternatives being presented on each trial, all consisting of ordinary objects with which each subje ct was highly familiar. The animal’s task was to select one of the two alternatives and place it with the sample. No differential feedback was given. Instead, the animal was praised for all its choices and given a small portion of a favorite food at the end of each test session. Six test sessions were run, each consisting of 24 trials. Sessions were given either daily or at 4%hour intervals. Two lists of 12 trials were constructed (see Table 1). Each trial consisted of a sample or target object, and two alternatives - one the taxonomic alternative, the other the thematic. The taxonomic alternative was always a token of the same type as the sample, different only in physical detail. For instance, the sample was a teaspoon, the taxonomic alternative a tablespoon; the sampiea large red apple, the taxonomic alternative a small green one; etc. The thematic alternative was chosen so as to exploit the known history of the animals. For instance, when shoe was the target, shoelace was the thematic alternative, shoes with laces being a common toy; likewise when lime was the target, knife was the thematic alternative since cutting pieces of fruit with a knife was one of the animals’ favorite play activities. The two lists (shown in Table 1) differed in only one respect. In list 1, all the target objects were ones for which the experimental animals (Walnut and Qpal) had been taught names; in list 2, they had not been taughtnames for any of the target objects. This difference had no bearing on either of the control animals, for Frieda had never been taught any names, and Sarah had largely forgotten those she was taught; at the time of the test, it was 13 years since she used any plastic words except for “same/different”.
Table 1.
Lists of test item
_________
.--
--
_-___I_-
Target object A~c~~~ives _---Knife Apple Pen Shoe CUP Carrot ?+dlet spoon Spray bottle DON Sponge scissors
Pen!paper Shoe&ace Cupljuice Carrot/k&e Mallet&rape Spmnkontainer
Sponge/waxer Scissors/paper
2. Clnnamed target objects
Target object Alternatives Paper Lid Coat Sock Nandywiper Lock &tray Jell0 Lime Dough Hose psrt Bubbles --
Hand
Lime/knife D 3-I Bu
ons were run, all of them with list I. The control rget object and the two alternatives without any ect. This served to establish how the individual sorted, r a taxonomic one. In a second test condition, anied by its name, that is, the name taught the stic word being placed alongside the target each target object was accompanied by a nt one for each target. A novel or nonsense he cIass of items that constituted words in the he class consisted of pieces of plastic, varying
Words
Table 2.
203
Test conditions and order in which they were given 1. Control. Sample: Target object alone 2. Sample: Target object + name of target object 3. Sample: Target object (named) + Nonsense word 4. Sample: Target object (unnamed) + Nonsense word 5. Control. Sample: Target object alone 6. Sample: Target object + “foreign” object
in shape, size, and color, each backed with metal so that it adhered to the magnetized language board. These items are nonsense or potential words in the sense that, though each conformed to the word class, they had never been used as a word. In the fourth test condition, each target object was accompanied by a “foreign” object, a different one with each target object. Foreign objects - for example, bottle caps, dominoes, bolts - did not cc>il’iirtilto the properties of the word class and thus could not be considered novel or nonsense words. List 1 was used in each of the four test conditions, and list 2 was used once in a special test condition (see Table 2 and results). The 12 tri:Js pf both list 1 and list 2 were presented twice in every test, the left-right position of the two alternatives being interchanged semi-randomly, giving a total of 24 trials per test. The order in which the test conditions was run is shown in Table 2; the control condition was rtrin tw~e, in the first and fifth position, to establish that the base condition was stable during the test period.
s
Figure 1 shows the number of taxonomic choices each animal made in each of the six tests. Consider first the two control tests: they show that all three young animals - Frieda, Walnut ard Opal - sorted on a thematic basis; whereas Sarah, the one adult an rn~l, sorted on a taxonomic basis. Sarah chose the taxonomic alternative 37 out of 48 trials; the three young animals chose it only an average of 9 out of 48 trials @ > .Ol). Language training did not appear to figure in this difference. Frieda, the nonlanguage-trained control, showed the same degree of thematic sorting as her two language-trained peers. Test 3, in which the target object was accompanied by its name, produced the expected effect. For the two experimental animals, Walnut and Opal, the plastic word that appeared alongside the target object was the name of the target object. Both of these animals came under the control of the plastic
mack
Figure 1.
0
I 2 8 1 CONTROL
3 I
4
5
6 lwkml
words: they shifted to the raxonomic alternative, choosing the object named untrained in words, was naturally unafto choose the thematchange, continuing to trend continued: Frieda (the young animal) chose the n all test conditions, eas Sarah (the older one) chose the taxonomic alternative on all test tions, the choiczs of the two animals being dete ined apparently by their respective ages. tests of greatest interest are the remaining three: 4, 5 and 6. The in t focuses on the two experimental animals, Walnut and Opal. Novel words affected these two young guage-trained chimpanzees in the same way as they do young children. both tests 4 and 5, when the novel or onsense words accompanied the target objects, both Walnut and Opal made redominantly taxonomic choices, an average of 20 out of 24 trials (p > .05). fact, by comparing tests 4 and 5 with test 3, we see that the experimental animals made as many taxonomic choices when given novel or nonsense words as when given the actual names of the target objects. Comparison of tests 4 and 5 adds the interesting fact that the effect of the novel word is felt whether the target object is named or not (target objects were named in test
Words
205
4, unnamed in 5). The marked effect of novel words on the experimental animals contrasts with the complete absence of effect on Sarah a Novel words had no effect on these control animals. Finally, test 6 confirms the legitimacy of speaking of the effect of specifically novel or nonsense words. Foreign objects, items that do not conform to the word class, do not produce a taxonomic shift; under this condition, both Walnut and Opal returned to the thematic alternative. This control, which incidentally is lacking in the child research, demonstrates that it is not simply the addition of any object (o*1 Dyuaen ment sound) that produces the shift, but only objects with wordlike properties. In brief? only animals that were both subadult and language-trained showed a shift effect, one that was as pronounced for novei words as for actual words, and one that did not occur at all for foreign objects.
The present results indicate that novel words lead to a thematic-taxonomic shift in the chimpanzee as they do in the child; the effect is not species specific. One can always maintain, of course, that, while children; and chimpanzees act the same, they do so for different reasons. This does not seem t& ideal occasion for such a claim, however, since there is no apparent reason for making it. On the other hand, the fact that the two species share the shift effect does 3ot wowu.tiauir #%?*Qi%l;C~ any deep correspondence beta-&en human and animal words. As suggested earlier in the Introduction, the shift effect does not tell us what a word is - even what a child (or ape) thinks it is - and there is reason to suspect that tlb * effect is a superficial one. For example, the shift effect is not only found in animals but in animals at an early stage of language training. Walnut and Opal had been taught 25 plastic words, primarily names of objects. The names of ~WG actions had only recently been introduced, and training in object-action pairs had just begun. Moreover, even this may be more than is needed; animals given may still show the shift effect. What is the minimal knowledge an individual must have to show the effect? We may find an answer to this question by noting that the effect of placing a novel word alongside an object is exactly like that of placing thti actual name of the object there. Both lea the individual to take the taxonomic alternative. For instarce, when Opal and Walnut were given a large knife as the target, a small knife and an apple as alternatives, both started by choosing the apple. Adding the plastic word “knife” - placing it alongside the target
- shifted both anim
;
they mw to
se, “knife’” is after
all the name of knife - yet this out@ than thematic associations; they exe titian with thematic associations, words response. ct; when placed alongside the target Novel words have exactly the ssociation, no less effectively than object, they too overcome the rds that are placed actual words. Thus, the animal appears to treat ct, no less so than alongside the tar bject as being the n y? Some previous data suggest an answer to this questhe actual name. tion. Sarah was taught many words with the instruction “X name of Y:t, X being a potential word (an unused piece of plastic), Y an unnamed object. But with the continual need for new words - every transfer test required several, an tests of this kind re given daily - the trainer found “X name of Y” tedious. She devised a si ler procedure. She took advantage of two containers of objects, one a 1 km of unnamed objects that were reserved as referents for new words, the other a supdv of appropriate pieces of plastic cut from long molds of plastic and back&i with a piece of metal so that the word would adhere to the magnetized writing board. Taking one of the not-yet named a3bjects in one hand, and one of the not-yet used words in the other, she held bath of them up f-VI the animal’s inspection. She then clapped the two tohenceforth found that words introduced in this manner could be name of thz object no less effectively than words introduced in a onal manner. a novel word and object together, on the one hand, and placing alongside the target object, on the other, appear to have the he animal treats the novel word as the “name of” the object in question. A naive chimpanzee would not do so, of course - only one that has been appropriately trained. n fact, the naive chimpanzee has great difficulty learning the first words ht it. Three hundred or more trials are often needed to introduce the animal to words. owever, once the animal has acquired first words, it acquires subsequent words in fewer trials, ultimately in a single trial. What does the animal learn that accounts for this transition? A series of tests showed that, in the course of learning its first words, the animal had learned to recognize members of the two classes - those of the word, on the one hand, and those of the referent, on the other. The animal could distinguish novel or unfamiliar members from nonmembers in both cases (all referents in this case were fruit and thus had class properties). Besides a knowledge of class properties, the animal had also learned the appropriate use of each class; for
example, to use plastic words to obtain Btot vice versa; to place words on the writing board, not in the m (Premack, 1976). At this early stage of training, astic words were virtually confined to a single use: to “request” fruit from
e trainer. More interesting language-like uses of the plastic words had yet to occur. For instance, plastic words were t yet used to answer questions, for example, “? color of apple” (What is color of apple?), “square ? cracker” (What is the relation between square cracker?), etc., which the animal “answered” by removin interrogative particle and replacing it with the appropriate word. But s the point - “advanced” language-like uses of the plastic words are not needed in order to produce either a shift effect, or one-trial learning. Both of these effects can be produced merely.by arranging for the individual to experience a co+ sistent functional, relation between the individual members of two different classes. If a member of one class can be used to obtain a member of the other class, and this holds for all members of both classes, the generalization leading from this experience will produce both one-trial learning and a shift effect. That is, this experience produced these effects iq11appropriate creatures, apes (and presumably children) being evidently appropriate. X can be used “to obtain” Y, though not equivalent to X is the “name of” Y, nonetheless produced a shift effect; we cannot therefore take the shift effect as evidence for the knowledge that a word is the “name of” a category. Note we are not arguing that the child’s experience was as narrow as that of the apes - restricted to the use of words as requests; certainly it was not. But children do use words to make requests, and since this narrow use was scfficient to produce a shift effect in the ape, we have no reason to suppose that it is not sufficient to produce the same effect in the child. This is the real problem: for all we know, the experience “X can be used to obtain Y” may be one of a small family of experiences that leads a child (or even an ape!) to treat “X and Y” as exemplars of the primitive relation “X name of Y”. But, of course, we cannot simply assume that - it is the most essential part of the problem we are trying to solve. Unfonrlnately, as should now be clear, the shift effect provides no leverage in solving that problem. Note that the effect of the novel word held 8s -:dell for target objects that were already named as it did for unnamed target objects (compare results of tests 5 and 6), casting doubt on another claim of Markman and Wachtel (1988), namely, that of “mutual exclusivity”. According to these writers, children acquire -words efficiently not only because of the assumption that words name categories (“taxonomic organization”), but also because of the assumption that if an object has o e name it will not have another (“mutual exclusivity”). However, if children hold this assumption, they hold it lightly,
for the assumption does
shift effect. For example, when d this is a “dax, find me another dax”, the shift effect induced is not confined to unn ed objects. Although “dax” may be the second: name the c ild is given for the object, it still produces the shift ees further resemble one another in that effect. Thus children an the use of unnamed objects. the shift effect does not hift effect really depends on the assumpe specifically eat@gories. ust the target object be an ti exemplar of a category? To be sure, in all tk amples at ha&. in the ape no less than in the child, target objects are of exactly this kind. Dolls, whistles, apples, knives, etc. are members of categories. However, that is exactly the problem: we lack tests in which the target object is not the member of a order to be certain that category membership is critical, tests are hich the target objects are not traditional members of categories, and novel words therefore do not name categories. Moreover, we must find that in such tests novel words do not lead to a shift effect. Since there no actual tests of this kind, we may fall back on a Gedankenexperimen ur Gedankenexperiment is exactly like all other shift effect tests except for one feature: the target object is not a category member. Instead, we use a picture of the child’s mother as the target object. As alternatives, we offer the standard taxonomic and thematic choices; for example, another somewhat different picture of the mother, and a picture of a familiar purse that the child easily recognizes as belonging to the mother. Let us se that the child shows thematic sorting in this case too, choosing the er’s purse rather than her picture. This assumption will enable us ito take e next step, and inquire about the effect of novel words. odeling our instructions as much as possible on those used by Markman we tell the child “Nina is what your mother was called when she was and ask her to “find another Nina”. Does this produce a taxonomic shift, leading the child to shift from picture of purse to that of f it does not, then all is well for the Markman et al. hypothesis: a” is not the name of a category, and therefore it should not produce a shift effect. On the other hand, if “Nina” has the usual effect of the novel word, then all is not well for the hypothesis. The child grasps, I assume, that “Nina”, as the proper (childhood) name of the mother, is not the name of a category (though probably the child’s understanding of this distinction merits study in its own right). In any case, in all tests of the shift effect only traditional category objects have been used as target objects, creating the imgression that the effect depends upon the knowledge that “words naive categories”. This is, I think, an unlikely assunlption and, in any case, an untested one. not
interfere
tit
the
words
209
What is a word? Information retrieval and reference Even a patient reader may now complain, “This article is ending without having made any progress on the principal question: What is a word? What is meant by ‘name of?” Let us consider two properties of words, one pointed out by behaviorism, the other by mentalism. First behaviorism: words are information-retrieval devices. Traditional discussions of language make the same point when they treat displacement as a hallmark of language. The ability to talk about things that are not present depends, of course, on the information-retrieval capacity of words. What does information-retrieval depend on? Mainly on a certain quality of mental representation. A species that suffered a major disparity between what it can perceive and what it can represent - a species in which representation did not do justice to its perception - could not talk effectively about what it cannot see. Words have no power as such; they borrow their power from the representations they retrieve. Were the chimpanzees’ plastic words effective information-retrieval devices? Dividing this question into two parts, l:re first established the nature of the chimpanzee’s mental representation, and then tested its ability to retrieve these representations with the plastic words. Match-to-sample tests were used in both cases. For example, the animal was given a taste of, say, peach, and then shown a green patch and a yellow one (or a paper silhouette of an apple and a peach, a peach pit and an apple seed, etc.). Three of the animals chose the correct alternative over 80% of the time; Sarah performed still better, successfully matching each of seven different features into which eight fruits were divided with every other feature (Premack, 1976,. The main point was established: there was no major disparity between perception and representation in the chimpanzee; representation did justice to perception. We then tested the ability of the plastic words to retrieve the information we knew the animals had stored. Plastic words were substituted for the former alternatives in the match-tosample tests. For instance, the animal was shown an apple stem as the sample., and offered as alternatives the plastic words “apple” and “peach”. With words as alternatives, three of the animals passed tests that they had failed earlier, when physical features were used as alternatives. In effect, the information retrieved by the plastic words exceeded that retrieved by actual parts or features of the fruit. This property of the plastic words will account for the animal’s ability to acquire words noncstensively. Sarah was taught the word “brown” with the instruction “brown color of chocolate”, the words “color of’ and “chocolate” being already established. Neither chocolate nor any brown object was pres-
ent. Next, she was given the instruction “take brown”, in the presence of four discs only one of which was brown, and chose correctly (Premack, 1976). ne can say of the information-retrieval side ef words that it is of little t does not seem quite for by conditioning. interest since it is we have already seen one accounts for it the right reaction. that a certain quality of mental representation is a major prerequisite - the sheer amount of information controlled by the Kxd is uncanny. Perhaps one can put this side of wor into perspective by considering what happens when opriately train chimpanzee is given the instruction: “X name of Y”. en so instructed, the animal’s ment presentation of the object Y is can be evoked mentally by X. red to the new word X. Hencefo s seems a remarkable transformation, an uncanny characteristic of mind. regrets to the readers for whom this property is demeaned by the fact that it can be accounted for by conditioning. mentalism we receive a second property of words, namely words ot all words but at least those that name objects. Here we could amara (1982), who treats reference as a special intentional state, one which young children can recognize. They can distinguish, he claims, when a speaker says “Betty” or “dog”. ~tc. w intention to refer to the item in question, and when, in contrast, the speak s simply musing aloud. Then our question becomes: is this special intentional state one that is recognized and/or instantiated by any nonhuman species? One way or another we need so we need to present both children and apes with speakers who employ the same words, in some cases with intent to refer, and in other cases without, and then determine whether either or both species can make the proper distinctions. Children, it goes without saying, will be far better at this discrimination than apes; indeed, there must be circumstances children understand nicely that apes cannot grasp at all. But are there no circumstances under which the ape can make this discrimination? That does not go without rtunately, there appears to be no direct evidence. The closest evidence concerns whether chimpanzees attribute any states of mind to others (Premack, 1988a; Premack & Dasser, in press; Premack & Woodruff, 1978). On the whole the evidence is positive, though there is ample indication that the animal’s competence is limited, that its “theory of mind” is not equivalent to the human one (Premack, 1988a). Nevertheless, if an individual attributes SOme kinds of intentions to others, as the chimpanzee apparently does, might it not then attribute specifically intention to refer? There is a chance. How similar is intentional reference to other kinds of intentions? We are, I think, beyond the polemical stage of this discussion; we see what some of the ques-
tions are and know how to test them (even as we know how to test for “theory of mind” in general). Children use words in ways that apes do not, one or more of which may concern reference. One that is especially suspicious is the strange reaction that children between approximately 10 and 13 months of age show to an object that is of special interest to them. For a period of perhaps two weeks the sight of the object produces an exceptional arousal state in the child. The arousing object in the case of a l(\.S-month-old male was a moving truck, a moving bus in the case of a 13.month-old male, and a swimming goldfish in the case of a 12-month-old fe ale. Each child pointed excitedly at the moving object, calling out the name repeatedly - %l”, 66bu!9S,“ef !” - all the while avidly seeking eye contact with its recipient. Sometimes the child had a favorite recipient, and endangered its life in an attempt to communicate with her, as when one child (riding in the front with its father) sought to climb out the window of a truck to reach its mother riding in the back. What was the meaning of the child’s behavior? Was it a request? Did the child want to reduce the distance between itself and the arousing object? Did it want to touch or at least see the object up close? One could test this more practically with the fish than with either the truck or bus. One took the fish bowl down from the shelf and offered it to the child. The effect was clear from the first test: the child stepped back, rejecting the opportunity to reach in and touch the fish, or even to draw near and observe it more closely. Her pointing arm dropped and she fell silent. Moreover, once the bowl was restored to its proper place the child immediately reverted to its former behavior - shouting (“ef! ef!“), pointing, and seeking eye contact. The child’s message was not that of a request for the object. This behavior had several other properties in all three children (although at least five children are needed for statistical significan , family size is rarely designed with statistics in mind). Each child behav in this manner with respect to only one object, for only a short period (perhaps 10 days), using a word which, though an early word, was not the child’s first word. For example, all three children used “mamma” at this stage. But no child behaved in this way with respect to the would-be referent here, that is, pointed at mother, shouted “mama!” while avidly seeking eye contact with father or some other recipient. Was “mama” merely a pleasing sound the child emitted in the presence of the appropriate party, while ‘Yu”, “bu” and “ef” were the real first words, that is, words that referred‘? This is a tantalizing question but not a good one because there is no apparent way to answer it. Finally, we see there is no incompatibility between the properties that behaviorism and mentalism emphasize. Any words an individual uses in acts of reference will, of course, be well associated with her mental representation
21
remtack
e objects to which she refers. trieval.
ce presupposes information re-
Bowerman, M. (1988). Inducing the latent structure of language. In F.S. Kessel (Ed.), The development of language and language reseurchers: Essays in honor ofRoger Brown. Hillsdale, NJ: Lawrence Erlbaum. Carey, S. (1985). Concepml change ila childhood. Cambridge, MA: MIT Press. Keil, F.C. (1983). Semantic influences and the acquisition of word meaning. Berlin: Springer-Verlag. Macnamara, J. (1982). Names for things. Cambridge, MA: MIT Press. Markman, E.M., & Hutchinson, J.E. (19s). Children’s sensitivity to constraints on word meaning: Taxonomic vs thematic relations. Cogrritive Pq&oiogy, 16, i-27. Markman, E.M., & Wachtel, G.F. (1988). Children’s use of mutual exclusivity to constrain the meanings of words. Cog&ive Ps@roiogy, 20, 121-157. Premack, 1). (1976). intelligence in ape and mun. Hillsdale, NJ: Lawrence Erlbaum. Premack. D. (1986). Guvagai? or the future history of the animal ianguage controversy. Cambridge, !:A: MIT Press. Premack. D. (198Sa). “Does the chimpanzee have a theory of mind?” revisited. In R.W. Byrne & A. Whiten (Eds.), Machiavellian intelligence. Oxford: Oxford University Press. Premack, D. (198Sb). Minds with and without language. In L. Weiskrantz (Ed.), Thought without language. Cambridge, UK: Cambridge University Press. Premack, D., & Dasser, V. (in press). Perceptual origins and conceptual evidence for theory of mind in apes and children. In A. Whiten (Ed.), Mindrecxding: The evolution, development and simulation of secondorder mental representations. Oxford: Basil Blackwell. Premack, D., & Woodruff, 6. (1978). Does the chimpanzee have a theory of mind? The Behavioral and Brain Sciences, I, 515-526.
Quine, W.V. (1960). Word and object. Cambridge: MIT Press. Skinner, B.F. (1957). iJJerbrrlbchbvior. New York: Appleton-Century-Crofts. Slobin, D.1. (1985). Crosslinguistic evidence for the language-making capacity. In D.I. Siobin (Ed.), The crosslinguistic study of language acquisition. Hillsdale, NJ: Lawrence Erlbaum. Smiley, S.S., 8 Brown, A.L. (1979). Conceptual preference for thematic or taxonomic relations: A nonmonotonic age trend from pre-school to aid age. Journal of Experimental Child Psychology, 28, 2@257.
Talmy, L. (1976). Semantic causative types. In M. Shibatami (Ed.), Syntax and semantics, Vol. 6: The grammar of causative constructions. New York: Academic Press. Talmy, L. (1983). How language structures space. In H.L. Pick Jr. & L.P. Accredobo (Eds.), Spatial orientution: Theory, research and application. New York: Plenum Press.