Cognition
v
Editorial Board Hioshi Azuma
Janet StojakCaplan
Janet Dean Fodor
Tokyo University, Hongo, Bunkyo-ku
Department of Psychology, Yale University, New Haven Conn. 06520, U.S.A.
Department of Linguistics, University of Connecticut, Storrs, Corm. 06268, U.S.A.
NoamChomsky
Paul Bertelson Lebomtok de Psychologie Exp&imentale,
Dept. Modern Languages and Dept. of Psychoiogy, Linguistics, MI. T., M.I. T. EIO-034, ambtidge, Mass. 02139, U.S.A. ambridge, Mass. 02139, U.S.A.
JerryFodor
ManfredBierwisch
&e ark Department of Linguistics, Stanford University, Stanford, Gdif. 94305, U.S.A.
Dept. of Psychology, Monash University, Ciayton, Vie. 3168, Australia
Ned Block of Philosophy, M.&T.,
Labonrtcfl y of Experimental Psychub?gy, Centre for Research on Percep tion and Cognition, University of Sussex, Brighton B&l, Ct. Britain
Faculty of Education,
Tokyo, Japan
t.Iniversit&Libre de Bruxellcs I1 7 Av. Adolphe Buy/, B-1050 Brussels,Belgium
Akademie der Wlssenschaften der DDR, Zent.&nstitut fir Sprachwissenschaft, Otto Nuschke Stmsse 22123 108 Berlin, G.D.A. Dept.
Cambridge, Mass. 02139, U.S.A. Melissa Bowerman
Psychoiogy Department,
University of Kansas, Lawrence, Kunsas 66044, U.S.A.
FranqoisBresson
Labonrtooln de Psychologie, 54 bvd. Raspail, F- 75006 &is, Frcmce
RogerBrown
Anne Cu’T:r
JamesE. Cutting
Psychology Department, Uris Hali, Cornell University, Ithaca, N. Y. 14853, U.S.A.
PeterD. Eimes
Walter S. Hunter Laboratory of Psychology, Brown University, Rovidence, R.I. 02912, U.S.A.
GunnarFant
Lab, of Speech Thnsmission, Royal Institute of Technology, S- IO044 Stockholm, 70, Sweden
Dept. of Aychokvy, Narvatxt University, C&s Fauconnier Cambridge, Mass. 02138, U.S.A. 9 Rue des Guillemites, 75004 hris, France mer
IL Bryant
Degrrrtment ‘of Experimental
Psychology,
DavidFay
Utdverslty of Oxford, South Aprk Road, OxfOn OXI 3UD. GL Britain
Bell Lubo~tc;rries, Warrenville - Napetville Road, Naperville, Ill. 60540, U.S.A.
DavidCaplan
IraFischler
Division of Neutw&gy, Ottawa CivfcHospital, Ottawa, Ont. KIS 2A3, Guta&
Department of 1”Jvcirology, Univemityof Florid&, Gainesville, i?l& 32611, USA.
KennethForster
MerrillCarrett
Department of Psychology, MS. T. El O-034, Cambridge, Mass. 02139, U.S.A. Lila Gleitman Graduate Sc41001of Education, University of Pennsylvania, 3700 Walnut Street, Philadelphia, Pa. 19104, U.S.A.
DavidT, Hakes,
Department of Psychology, Ur,iversity of Texas, Austin, Tex. 78712, U.S.A.
HenryHecaen Directeur d’Etudes,
Ecole Pratique des Sautes Etudes, Unitt!de Recherches Neuropsychologiques. I.N.S.E.R.M., 2, rue d’Alds& F- 75014, Paris, Fraace
Michollmbert
Labora toire de Neuro.
physlologie, CollLge de France, 1I Place Marcelin BertheYot, F. 7.5005 Paris, France
B&kJel Inheldcr
Fact&t! de Psychologie et des Sciences de l’&iuc&& Univ&tt! de GenCve, Cl&I 211 Geneva 14, Swiaerland
Marc Jeannerod Laboratoired ? iG?urupsychologie .Experimentale, I6 Av. Doyen L&pine, F-69500 Bran, France
Willem Levelt Max Planck?nstitutfur Psycholinguistik, Nimegen, The Netherhmds
John Lyorls PhilipJohnson-Laual Dept. of Linguistics, Laboratoryof Ex perimen tal Adam Ferguson Building, Psychology, Edinburgh EH8 9LL. Gt. Britain Centre for Research on Perception and Cognition, David McNeil1 Sussex University, Brighton R>Nl9QG. Ct. Britain Departmentof Behavioral Sciences, Committeeon Cognitionand Peter W. Susczyk Communication, Dept. of Psychology, Universityof Chicago, Universityof Oregon. 5848 South UniversityAvenue, Eugene, Oreg. 9 7403, Chicago,Bl. 60637, tt S.A. USA. Jerrold 1. Katz Dept. of Linguistics, CUNY GraduateCenter, 23 W42nd Street, New York, N.Y. 40036, U.S.A. Mary-Louise Kean CognitiveScience hogram, School of Sock?1Scienses, Universityof Califomio, Irvine, Calif. 92717, U.S.A. Frank Keil PsychologyDepartment, Cornell University, Ithaca, N. Y. .14853, U.S.A. Edward fUima Dept. of Linguistics,La JO&I, Universityof Califomlb, San Diego, Cz1i.f92037, U.S.A. Stephen hf. Kosslyn
Department of Psychologyand
So&d Relations, HarvardUniversity, Willhzm James Hall, 33 KirkkandStreet, ambridge, Mass,0.2138, USA.
HarlanLane Department of Psychology, Northeastern Unive&y, 360 HuntingtonAvenue, Boston, Mass,02115, US.A.
John Marshall Neuropsychology Unit, Radcliffe Infirmary, WoodstockRoad, Oxford OX2 6HE, Ct. Britain William Marslen-Wilson Max Planckhstitut fir Psycholinguistik, Berg en Dalseweg79, Nijmegen, The Netherlands Jod Morais Laboratoue de Psychrologie Experimentale, UniversiteLibre de .Bmxelles, 117 Avenue Adolphe Buyl, B-l OS0Brussels,Belgium
Michael Posner Dept. of Psychology, Universityof Oregon, Eugene, Ore. 97403, U.S.A. Dnvid Premack PsychologyDepartment, Universityof Pennsylvania, 3813.15 WalnutStreet, PhEadelphh, Pa. l 9174, U.S.A. Zenon Pylyshyn Dept. of Psychology, The Universityof Western Ontario, London 72, Or&, CImada Audr8 Roth Lecours Hotel-Dieude Montreal, 3840 rue St. Urban, Montreal,Quebec H2W 1 T8, Canada Steven Rose Bic-!ogy Department, The Open University, WaltonHall, MiltonKeynes MK 7 6AA. Ct. Britain Scania de SchBnen Laboratoirede Psychologie, 54 BoulevardRaspail, 75270 P&-isCddex 06, France Tim Shallice MRC Applied Psychology Unit, 1S Chaucer Road, CambridgeCB2 2EF, Ct. Britain
Dan I. Slobin John Morton MRC Applied Psychology Unit, ~~~~rree~$,?$~~~~gy# 15 Chaucer Road, CambridgecB~ 2EF, Gt. Britain t?erkeley,C”!i$ 94720, ‘U.S.A. George Noi,:et Laboratok de pSycholog& 28 rue Serpente, 7.5006 Parts,France
Elizabeth Spelke
Psychology Department,
Universityof Pennsylvania, 381.5 WalnutStreet, Phibdelphia, 84.19104, U.S.A.
Daniel Osherson Mark Steedman 2OC-124 (DSRE), Departmentof Psychology, M.I.T., Universityof Warwick, Gzmbndge, Mass02139, U.S.A. Coventry CV4 7AL, Ct. Britain
Sidney Strauss; Department of Educational Sciences, Tel Aviv University, Ramat Aviv, Israel
EdwardWalker M.I. T. Center for C+gnitive
Science, 77 MassachusettsAvenue,
Cambridge,Mass.02139, U.S.A.
Michael Studdert-Kennedy Department of Communication Arts and Sciences, Peter Wason @ueens College, Psycholinguistics, Qty Universityof New York, UniversityCollege London, Flushing, N. Y. 11361, U.S.A. Research Unit, David Swinney 4 Stephenson Way, Department of Psychology, London NW1 2HE. Ct. Britain YldftsUniversity, Medford, Mass.02155, U.S.A. Virginia Valian 221 I Brvadwdy, New York, N. Y. 10024, U.S.A.
Ken Wexler School of Social Sciences, Universityof California, Irvine, Galif ?2 717, U.S.A.
Deirdre Wilson Deprrtment of Phonetics& Linguistics, UniversityCollege London, Gvwer Street, London WC IE FisT, Ct. Britain Edgar Zurif AphasiaResearch Center, Boston UniversityMedical Center, I50 South HuntingtvnA venue, Room ClS-5, Boston, Mass.02130, U.S.4. ‘iermina Sinclair de Zwart Centre d ‘Bpistemvlogie Gt%tPtique, Universityde Genkve, CH-121I Geneva, Switzerland
Cognition,10(1981)l-5 @Elsevier Sequoia S.A., Lmsanne - Printed in The Netherlands
1
Editorial JACQUES MEHLER SUSANA FRANCK
It was almost by mistake that we noticed some months ago that Cognition was ten years old And with journals as with children, growing has both the advantages of experience and the (dangers of stultification. Thus, once we had resigned ourselves to the fact that we actually were coming of age, we had to fend off aging and stiffening structures. As a result, two decisions were rapidly taken. The first, and more administrative of the two, was to increase the journal’s periodicity from one to two volumes per year. The second, and mare creative,. was to ask several colleagues to write a few pages about their work and how they looked on it given the state of cognitive psychology as Rwhole, and what they thought developments in the discipline would be like in the coming years. These two editorial decisions can in some ways be looked on as controversial and a few comments seem therefore to be in order. We only decided to ex.pand into two volumes per year after a great deal of hesitation. Indeed, although the number of journals in the domain has increased over the last ten years, we are not convinced that the quality of the work in the area has met the promise it displayed a decade ago. Undeniably, considerable progress has been made on some formal fronts both in linguistics and AI. In cognit,ive psychology, however, progress is less obvious. A number of new paradigms have become accepted working tools and some new fields have opened up. But in contrast with, say, molecular biology over the last twenty or thirty years, one certainly does not get the impression that any revolution has occurred in our field. Experiments have perhaps increased in ti&.istication and a few optimists ncj doubt believe that some major development is just around the comer but m the meantime little has really changed. Thus, as journal editors waiting for the supreme breakt-hrough we had the choice of becoming very tough minded and selective in order to keep the number of pages published in Cognitim down to a minimum or the option of increasing the size of the journal while preserving its quality. The decision was not an easy one and it is precisely because no ma,ior new discovery has as yet shaken the field that it seemed wise to settle for the second course of action. Indeed, the exploratory nature of our work makes diversity enormous, areas of interest numerous and polemics plentiful. For the time being, then, since all the material swept up by the cognitive tidal wave could be of potential interest to the field as a whole,
2
Editorial
we decided to increase the number of printed pages available in the journal in the hopes of broadening its scope and attracting more interesting contributions while still continuing to reflect the best of the field and, to some extent, shaping its form
Editorial
3
If the disappearance of the more ideologically oriented part of the journal has been a source of some disappointment to us, we must confess frustration with the uniformity of the submissions we tend to receive. Of course, a journal like Cognition is in a way the hostage of progress in scientific research while at the same time (providing it is successful) it influences the orientation and form of the work produced. Indeed, to a large extent, contributions shape a field and journals perpetuate the image practitioners have of their area. The process is self-fulfilling, stifling to any attempts at originality and very difficult to break at any level. It was largely in discussing this problem that we hit upon the second way of celebrating Cognition’s tenth anniversary. As we mentioned earlier, our second decadal move was to ask several colleagues for statementq about their work, the future of their endeavors, their evaluation of the field and, their views on how their goals might be attained. Although our invitation letter was admittedly very vague, the aim should have been quite clear. We were inviting authors to say what they pleased without any constraints other than a page limitation. They were to have the opportunity to make whatever statement they wanted in ways that are not habitually available to scientists. It was our hope that the outcome of this unusual enterprise would somehow reflect the state of cognitive psychology today as well as its potential for development in the next decade or so. Unfortunately, as we all know this kind of futurology is largely based on myths. Indeed if we could actually predict today what is going to happen in basic research in ten years time we could save a lot of energy and take a short cut up to the investigations that should be taking place then. But many of us do sit on committees or pariicipate in administration of one kind or another and are frequently asked to provide opinions, often vague or even thoroughly confused ones, as to what the shape of ths field will be in the coming years. Thus, in spite of the inherent difficulty in prospective thinking we are all called on to engage in it at one time or another. Indeed, computers are a useful case in point. When they were first introduced, all the psychological laboratories in the richer countries where the discipline was more or less well represented were rapidly equipped with them, Whether we care to admit it or not, computer fever overtook us before we had had the time to think about where it idlight carry us. In fact, although we all agree that computers are excellent support tools, our degree of confidence in their usefulness at furthering advances in the field in significant ways is by no means homogeneous. Indeed some would say that ‘...nowdays we are becoming aware that the infinite potential of electronic psychology was a senseless proposition because the potential was mainly on the surface
4
Editoricl
behavioral level rather than at the conceptuallevel.This latter levelbecamemore and more impoverished in the midst of programsL nd simulationsgeneratedby the new wave. And as a matter of fact, many psychologists becameawarethat it wasas easy as it was useless to produce new proceduresand simulationsif they did not stem from a true theoretical psychological insight.Thcs it is even possible to claim that computers have drawn psychologis?s away from a truly perspective stance where technical, behavioral ‘and theoretical aspects are considered in conjunction with social and ideological ones’. Interestingly enough, this last sentence is a gloss from I. Berio (1981) that we have translated rather freely taking the liberty of modifying by substituting the word psychology for music and behavior for acoustics. Whether or not we agree with the above citation our point in using it is that there are currently a number of people making similar statements in areas as diverse as urbanism, sociology, psychology or the arts. Thus, to some exteirt, it must be true that the pressures exerted by the environment are homogeneously strong regardless of what the aims or needs of the various fields may be. This kind of determinism is, .)fcourse, to be expected but whether it should be accepted without any criical evaluation is another matter. In a field like ours, the only way in which we can understand whether developments should be attributed to fashion incorporating outside pressures rather than to true theoretical developments is through discussion between the people partici -3atingin the elaboration of the discipline. For this reason we thought that it would be exciting to have a number of investigators of varying ages speculate about the field and its future. Indeed, a utimber of trends do emerge from the manuscripts appearing in this special issue. There are those who are convinced that information processing and artificial intelligence hold the keys to the future; those who believe that the1 is no future for the field as a whole and those who believe that cognitive science will only come of age when it has developed a more descriptive approach displaying greater coIlcern for ecological parameters. Finally, there are those who have refuseJ to engage in any speculation whatsoever and have opted to remain close to the data they have gathered within the paradigm they have developed. But regardles: of the nature of the contributions we have received, one feeling pervades our reading of all the articles and that is that the field is in rapid expansion without any homogeny of approach. Interestingly enough, moreover, we get the impression that there is some correlation between narrowness of perspective and the difficulties encountered by the younger writers in gaining recognition by the community. If this is true, the trend may well reflect a situation that we already ad.dressed in an Editorial written five years ago (see Cognition, 4, 1).
Editorial
5
It is obvious that although computers are going to play an increasingly important role, the neurosciences will also be occupying a dominant position. In the first place we stand to inherit a number of the tools and procedures used in disciplines related to the neurosciences which will allow us to carry out our empirical tests in more carefully controlled situations. In the second, the neurosciences will hopefully yield some explanatory devices that will help us to account for growth and evolution of cognitive capacities. Indeed there are few, if any, terms in cognitive psychology to account for the dynamics of the behavior noted and it appears probable that any causal account of development will be couched in terms that are more biological than psychological. Nonetheless, we will also have to resist the temptation of reduc.ing phenomena to neurological terms that should only be accounted for in tenr:s of processing.. Luckily, a number of authors already seem aware of the dangers that reductionism holds for psychology. Last, but by no means least, cognitive psychology will have to remain open to considerations about the well-being of mankind and the needs of society as a whole. Wowever, as is the case in all burgeoning areas the options must be left open to incorporate the unexpected developments that none of us can predict for the coming years. Finally, we should like to close this statement by referring to the journal itself and the key role that has been played by our referees over the years. We should like to thank all the present and past members of the Editorial Board as well as the many anonymous reviewers who have generously provided us with their opinions and helped colleagues with their comments and advice. The time and energy you have invzsted in Cogrzition and your implication in the life of this journal are what keep it going. It is our hope that we will be able to continue to call on you in the future and that our circle of active collaborators will increase steadily. In addition, we hope that the readers of Cognition will also help us to make the publication dev&p into an instrument that over and above representing the most excellent work in the field, will help our research and thinking to gain in perspective.
References Bedo, L. (1’ 1) htervistu Sullo Musica a cum di Rossana Delmonte. Guis. Laterza e Figli Spa, Roma-
Bari. Mehler, J., Bever, T., and Franck, S. (1976) Editorial, Cog., 4,
pp.7-l
1.
Cognition, 10 (1981) 7-15 @Elsetier Sequoia S.A., Iausanne - Printed in The Netherlands
7
Time for a purge TONY AlIES* Brighton, England
What is cognitive psychology? The question had never concerned me until ten years ago, just about the time Cognition was born, it suddenly seemed that there were surprisingly many cognitive psychologists. My own vision of cognitive psychology had been rooted in language. And judging from the preponderance of language-related research in this journal, thz same view must , :hared by most readers and writers of Cognition. The great attraction of language to me as a psychologist is that we know it is real. Its structure is subtle and systematic: it is a natural phenomenon. This is not the case for free recall, for example. People can remember telephone numbers, but no one would claim that this is natural in the way that speaking and listening are. By saying this I mean to do more than express my personal interest::. The point of natural science is to take a naturally occurring phenomenon under some characterisation and then try to understand it. Language presents itself, therefore, as an ideal object of e+udy. The subject matter of much of what is called cognitive psychology, however, is quite different. My week of discovery, ten years ago, began with a glance at Neisser’s classic Cognitive Psychology (1967). I swore I would never read beyond the title page. The contents belonged to a field that had long since settled in my mind as ‘complex reaction times’. After looking at the matter further, irritation turned to alarm. It seemed that other leading figures in ‘cognitive psychology’, many heading departments where I would one day have to seek work1 were no more than semi-reconstructed behaviorists from the verbal leaning and verbal behavior underworld, from ‘concept formation’, even from the gloomy abyss where de-natured rats and pigeons lived out their wretched lives. Others were long-practiced at deriving learning curves from Markov models of memory. . Under heaven-knows-what reinforcement schedule all had evidently learned that the addition of words like ‘semantic’ to their repertoire of publication titles would give them easy entry into a new and expanding field.
“Reprint requestsshould be addressedto A. Ades, 19 MontpelierPlace, Brighton,BNl 3BF, East rdswx, England.
8
Tony Ades
of the theses of this paper is that the substantive content of most of what passes as cognitive psychology is predictable from this historical concern with less-than-natural phenomena. A second theme will be the social realities of the structure of the profession and of research practice; these factors determine the progress of the research. I will draw particular attention to the system of promotion, and to the role of computer science. There are, of course, several strands of research that have remained oriented to language and thought as natural processes. Transformational lin and its subsequent developments is pppeminent. However bizarre its formulations become, the commitment to language as a natural phenomenon in need of explanation is inherent in its (structuralist) method. The new linguistics has inspired various schools of experimental psycholinguistics, such as those at MIT and Austin, Texas, where the flavor of research into language acquisition, comprehension and production has remained essentially unchanged over the last decade. These are among the strands of cognitive psychology with which Cognition has been concerned. But it is not my intention here to evaluate this work or try to guess its future. My intention is to highlight the other cognitive psychology, which is essentially parasitic, to provide perhaps an outline of a field-guide by which the two can be distinguished, and to understand the relationship between them. One
How to spot non-findings, non-phenomena, and dis-simulations Here is one of my favourite non-findings. Subjects were given the names of two animals, say tiger and horse, and asked to rate how similar they were. This was done for a number of names in all possible combinations. The results were fed into a factor analysis, and hey presto--three factors ‘emerged’ as covering most of the variance. These appeared to be size, ferocity and human-ness. The conclusion is that the semantic structure of animal CORcepts is a vector in this 3dimensional space. Like so many experiments, this one is not actually about anything that people ever do or think. Nobody could believe the conclusion. (Dissenting reader!: are respectfully directed to footnote’ ). What in fact emerges is that ‘If animals=n be describedin termsof this gdimensional space,how arewe supposedto distinguish between them on other dimensions like numberof stripes,numberof knees, etc. If there are other dimensions, will we not end with as many dimensions as there are ways of talkingabout animals? Eventually,when the experimentis extended to other groupsof words,will we not have to postulate as many dimensionsas therearewordsin the language? If dissentingreadersstill dissent,they shouldprobablyturn to otherarticlesin this volume,if they havenot alreadydone so.
Tim for a purge
9
such expe~n~ents a very tractable, and very do-able. The factor analysis even
provides an interp This is what is fact) ‘work’. Pesp (Certa%nly, if they do no periment however, the )
riment, in a nutshell, ‘works’. sometimes (far too often, in ist believing that progress is being made. y will not be published). Besides the exthat invariably works is the computer
knows that for people who are trying to build computer models, to refrain from addin s to the program to make it ‘work’ is more than most mortals can do. Even more obvious is the distinction between an effort to understand a natural phenomenon and a program that mimics without trying to model the natural processes themselves. What is less to be expected is that psychologists should take thePe programs seriously as insights into the workings of the human mind. A recent example is the acclaim which has greeted the Harpy speech recognition program (see Klatt (I 977) for references). This program was funded by the Advanced Research Projects Agency of the US Department of Defense, and designed to meet a number of purely engineering criteria: accepting connected speech from several speakers, with moderate speaker tuning, in a quiet room, with a 1000 word vocabulary, with an artificial syntax, in a few times real time on a 100 MIPS machine. Harpy ‘worked’. Of course, Harpy was not supposed to have any likeness to human speech perception. Yet Norman (1980) writes ‘The relative success of Harpy . . . comes as a surprise to many people.‘; and ‘Its performance is good enough for a number of eminent people to take it seriously’ (p. 394). Norman, who is evaluating the program as a psychologist not an engineer, does not tell us what is ‘good’ about it. As for its ‘success’, it appears that all there is to say is that it ‘works’. Concepts like ‘phonologically .well-formed’, or syntax, can apparently be dispensed with. I have given two examples: an experiment on a non-phenomenon, and a psychologist’s reaction to a computer program that simulates, or more accurately, dis-simulates, speech recognition. It is for readers to judge whether these are representative of the bulk of what is called cognitive psychology. There are without doubt quite reasonable research enterprises, too, many of them in the pages of this journal. But it is also true that the list of w-:11understood non-phenomena is large and long, and that computer dis-simulations of the most hopeless kind get wheeled out and discussed ud nauseam. Why so many non-phenomena?
T’hiqis not the first time that the state of the field ins been deplored. Newell’s paper ‘You can’t play 20 questions with nature and win’ (1973)
10
Tony Ades
diagnosed an obsession with the accumulation of facts, and a mindless adherence to well-worn contrasts like serial versus parallel and continuous verssus all-or-none. (Newc?ll’s list of 59 phenomena reveals, on closer inspection, a very high .percentage of non-phenomena). What is surprising is that Newell, and Simon (1980) after him, underscore the symptoms so well without making any effort to identify the causes of the disease. One possibility is that natural phenomena are more complex than the nonphenomena beloved of experimental psychology. The lure of the nonphenomenon lies in its tractability. One has a sense that natural phenomena are, relatively speaking, either understood or not: progress is by sudden leaps of insight. Whether or not this is true, the belief that real phenomena are too difficult to study is wides;pread. ‘Eariy in the history of psychology’, explains Feigenbaum (1963, p. 298), ‘the psychologist invented an experiment to simplify the study of human verbal learning.’ Human verbal learning was here conceptualized as the rote memorisation of nonsense syllables. The implication is that what children spontaneously do to learn their native language is too difficult to tackle. Of course, those were early days. Yet Simon (1980) specifically singles out Feigenbaum’s Elementary Perceiver and Memorizer model as exemplifying a solution to Newell’s dilemma*. But it is not enough to identify the lure of the tractable as the culprit. We
must ask why it is that cognitive psychology is not a pursuit conducted by reflection in a comfortable armchair, with plenty of paper and pencils, directed at the Innei Universe. One obvious factor is behaviorism. For a classic behaviorist it makes no sense to distinguish between phenomena and non-phenomena, Both are reducible to the same components. No ‘task’ can be any more natural or reprosentative of real human performance than any other, simply because we are blank slates upon whom anything can be written. If it were to be universally accepted that there is, for example, a crucial difference between human linguistic performance and th’e performances represented in Newell’s (non)phenomena, in t,:rms of their ability to tap actual mental processes, then the field would bs. in much hea’thier shape. There would also be far fewer publications. At the risk of being tasteless, the pressure to publish, which only very few scientists can ever escape, makes aSimon offersEPAhias a solutionbecauseit constitutesa singlemodeluf an exhaustivelyexplored task. These were among the criteriathat Newellhad set. Simonpointsout that an additional benefit of EPAM is that it is quantitative.ButhL assertionthat ‘mod&shouldgive prominence to the quantitative invariantsof the human informationprocessingsystem’(p. 537) fligs in the face of all expetiena with naturalIanwage, where informationtheoretic approximationswem a notorious flop, and wherelinguisticsseemsto have mademoderateprogresswithout quantifyinganythingat all.
Time for f4 purge
it
11
all but imp0 ible to ever give up the endless treadmill of facts and weary
ions of which Newell complain Much more is involved than the opp trad first in print. Grants, egrees, promotion all depend very largely on the sheer number of publications. Under these conditions of chaotic haste and competition, how can anythin be expected but close adherence to the tractable experiment, the contrast with one more parameter varied? Why so much corn utet dis&n~ulatisn? The behaviorist tradition again has much to answer rbr. Remember that all that can be claimed for programs like those spawned by the ARPA project is that they start with speech sounds and end with some form of phonetic or lexical representation. It is threir success in effecting this, not the plausibility of the intervening steps, that arouses the interest of psychologists, Similarly, programs that start with word strings and end with parsings have been recommended, not for the insightfulness of the linguistic analysis implied, but simply because they ran. It is, of course, the distinguishing mark of behaviorism in its strongest form that psychology should discover the stimulus-response relationships and ignore intermediate variables: the theoretical construct has no validity. No one, of course, believes in this sort of thing any more. (?). But the tradition lives on. The rise of computer science has brought into play an entirely new set of forces, which also give impetus to dis-simulation models. As is frequently observed, in every field it touches, engineering, management, medicine, teaching or the office, computer science is introducing new types of fragmentation of work, and new demarcations of duties. Cognitive psychology is no exception. Computer scientists arc trained to use and devise programming compilation, etc. Psychologists, techniques for memory, pattern-mat&in by contrast are not easily able either to create these modules or build complex systems from them. Simply because computers are used to perform typically human functions, computer science techniques quickly become working models of various mental processes. To be ‘quiteblunt, psychologists have always been relatively unimaginative in modelling complex systems rigourously, .VIthey borrow from computer science. If this seem:+t’ar-fetched,one need only look at what has ‘ztually happened. Twenty years of experimental work on speech perception gave us endless studies of categorical perception, hemispheric lateralisation, dichotic lag effects, the speech mode (all prime non-phenomena),and absolutely nothing that told us how people understood fluent speech. Hardly surprising that
12
Tony Ades
engineers filled the vacuum with engineering solutions. Hardly surprising, too, that psychologists take these solutions so seriously: there are no other solutions to talk about. Cognitive cuckoo
What we are witnessing, then, is the displacement of cognitive psychologists from their field of study. (I refer, of course, to those few who have actually been in their field of study). The field falls to ‘cognitive science’, and the inside front cover of Representation and Understanding: Studies in Cognitive Science (Bobrow and Collins, 1975) reveals its cuckoo-like role: “This book contains discussions on the problems o!’ constructing and modelling intelligent systems by workers in the field of cognitive psychology, linguist&s and artificial intelligence. It deals with how knowledge can be represented, how _mcm:ory structures an intelligent view of the world, and how memories for stories, paragraphs and plans can be structured.”
The wording has been chosen with diabolical ingenuity: it neither confirms nor denies that the book is concerned with the study of natural phenomena. The contents, however, leave little doubt. ‘The strongest support for procedural representation comes from the fact that it works’ (Winograd, p. 190); ‘the best representation for a body of knowledge depends on how that knowledge is to be used by the program’ (Bobrow and Brown, p.104); ‘Our two goals for SOPHIE’s natural language processor are efficiency and fiiendliness’ (Brown and Burton, p. 330). The poiut here is not simply the distasteful nature of the research, but the fact ,that it is represented as something that cognitive psychologists would be interested in. In fact, the authors quoted above are to be welcomed for their frankness: other contributors (for example, W jods, Rumelhart, Schank, Abelson, Nash-Webber) make passing references to human ability, but then procede to construct models whose connection with human performance is less than tenuous. In terms of the research strategy one would adopt, there cannot be any question that understanding the hum&n mind and getting a machine to do something that humans do are radically different. Yet one finds over and over that the two are practically interchangeabb. The boundary, which shouldbe so clear, $6deliberately made fuzzy. One can only conclude that this collapse of consciousness in psychology, and the resulting Miltration by computer technofogies, has been made possible by the lingering infection of behaviorism.
Time for a purge
13
The driving force behind the new non-synthesis, however, is quite a different matter3. The commercial pressure to replace intelligent living labor with the dead is immense, and growing. Teachers, engineers and office workers made redundant by computer aided instruction, computer aided design and the automated office, can look forward to whiling the hours away with microprocessor chess masters. Cognitive science is bringing us, too, military and internal security applications in the form of ‘smart’ bombs, data bases of various sorts, and telephone surveillance, These consideration:; have played a crucial role in the deverrjpment of many computer technologes--including, of course, compu;,srs themselves. How can we expect the penetration of cognitive psychology by computer science to progress in the future? At a trivial level the continued increase in the speed of computation will mean that many programs that did not ‘work’ before, suddenly will ‘work’. Psychologists will be confronted with many more dis-simulations of human conceptual processes. One might hope that 3It is extremely difficult to understand the real nature of this ‘non-synthesis’ of understanding the mind and making a rrachine mimic it. Neither behaviourism nor economic and political pressures satisfactorily explain it. I have <*alledit a ‘new’ nonsynthesis only because its present character seems subtly different from earlier manifestations. Taking the Bobrow and Collins volume as a reference point, it is worth looking at previous collections. Superficially, Compufers and Thought (Feigenbaum and Feldman 1963) presents a similar mixture of pure science and pure technology. Chess-playing programs, theorem-provers, and pattern-recognisers lie alongside a heuristic for assembly line balancing, a simulation ot’ investment behaviour, and a discussion of what the Russians may be doing. (The involvement of the military-industrial complex, which has remained a constant throughout the history of the (non)-synthesis, is well-documented on pages ix-xii). What is different, however, is that the chapters on human cognitive processes make a serious effort to compare human performance, plus the theoretical constructs required to explain it, to the detailed logic of the computer model. Reldting them only at the input/output level, which is the current practice, was transcended, and a genuine synthesis achieved. Examples are General Problem Solver (Newell and Simon}, EPAM (Feigenbaum), concept formation (Hunt and Hovland) and pattern recognition (Uhr and Vossler). Whether or not these efforts were misguided in the light of present knowledge is, of course, beside the point: it is the thought that counts. Stepping still further back, ta the coaference proceedings Mechanisation of Thought Recesses (Blake and Uttley 1959), or to some of the writings of Warren McCulloch (for example ‘Machines that Think and Want’, written in 1950), it is clear that astonishing levels of synthesis were being considered. This was not necessarily to be a convergence of human psychology and artificial intelligence as *NCnow know it, In that period theoretical neurophysiology, the study of neural net:+ was more visible. The bchpviour of neural nets was being compared to the behaviour of parts ot the nervous syster I (its physiology) and to the behaviour of the whole organism (its psychology). At :he ‘ame time the rrathematica! construction of the nets was compared to the fine structure of the nervous system itself (for example, Shall and Uttley 19c3). Again, these efforts may have been doomed at the outset by assumptions we now know are false. Nevertheless, the brilliance of the conception remains. Why has there been a progressive deterioration of relations between com+ters and psychology over i.he last 30 years? I shall leave the reader with the following observations. 30 years ago models were mathematical creatures, not actually embodied in a machine. In the sixties, several of the most well-known programs were in fact ‘hand-simulated’. Now, programming languages are so powerful that one can try out ideas (to see if they ‘work’) at the terminal, without ever hawing thought about them before.
14
Tony Ades
they will cease to take them seriously, as they can be expected to become more and more implausible. Unfortunately, the past decade has shown that it is only those areas which have retained a sense of identification w-lib natural processes, particularly with language, that remain relatively undistorted. And even there, the huge funding programme from the Sloan Foundation is aimed precisely at ‘unifying’ cognitive psychology, linguistics, and artificial intelligence. The pure and the applied Ii1 trying to understand why cognitive psychology has become so riddled with non-phenomena and dis-simulation, one possibility is that there may be a tendency among scientists doing applied work to legitimize or glamorize their research by preseming it as ‘pure’. In this final section, I shall look at this question from the other side. How do we understand the ‘pure’ scientists’ attitude to applied research. A particularly scandalous example is neuropsychology/neurolinguistics, where the last ten years of ‘pure’ research has provided absolutely nothing for the hospital populations who serve as its experimental guinea pigs. This is the admission of all the neuropsychological and medical op:nion I have canvassed -in spite of the frequent plea that the work would lead to useful diagnostic tests. It seems that research suited to the needs of these patients, their relatives, and the health care workers responsible for them merits no atteniion. Neuropsvchology, which is now covered by several specialist journals, makes it clear that intellectual challenges are valued more highly than social and human ones. It is work with ‘theoretical’ appeal that will bring publication, promotion, approval. Work with social value (which is, in fact, just as intellectually demanding) is simply not considered. I present this as an objective fact- one that needs to be understood. Neuropsychology could, after all, be an applied field. Indeed, the paradox deepens at this point; for when we consider the theoretical results that have emerged, there seems to have been little progress since Wernicke. The field is rife with opportunism, the pursuit of non-phenomena, the hopeless effort to understand the disruption of mental functions without trying to characterize them in their non-disrupted form. The patients are for the most part used to provide gratuitous confirmation for theories developed in neighbouring fields. I invite readers and writers of Cognit#onto discuss with their colleagues their rttitude towards applied research, and to the structures in the practice of science which the example of neuropsychology brings to light.
Timefor IXpurge
15
Happy Birthday, International Journal of Cognitive Psychology The picture I have painted is not pretty; but it is not a pictule of cognitive psychology as I understand it. The living ghosts of behaviorism that haunt the laboratories, the pressure to publish, the computer boom: these are some of the factors that shape the field. It is time for a purge. Cognition has proved itself unique in both upholding the field as a natural science and in occasionally takrng up relevant sociai and political issues. I hope that in the coming decade it will open its pages in a more deliberate way to the kinds of questions I have tried !o raise.
References Blake, D. V., and Uttley, A. M. teds.) (1959) Proceedings of rhe Symposium on Mechanizationof Thought Processes,National Physical Laboratory. HMSO, London. Bobrow, D. G., and Collins, A. (eds.) (1975) Representationand Understanding:Sxdies in Cognitive Science. New York, Academic Press. Feigenbaum, E. A. (1963) The simulation of verbal learning behaviour. In E. A. Feigenbaum and J. Feldman (cds.), Computers und Thvtght. New York, McGraw-Hill. Klatt, D. H. (1977) Review of the ARPA Speech Understanding Project. J. Acoust. Sot. Amer., 62, 1345-1366. McCulloch, W. S. (1965) Embodimerstsof Mind. Cambridge, MIT Press. Neisser, U., (1967) CognitivePs.tichology.New York, Appleton Century Crofts. Norman, D. A., 1980. Copycat science or Does the mind really work by table look-up? In R. A. Cole (ed.) “erceptionand Reduction of Fluent Speech. Hihsdale, NJ, Fsrhaum. Newell, A. (.,73) You :an’t play 20 questions with nature and win. In W. G. Chase (ed.), YisualInformation plocessing New York, Academic Press. Shall, D. A., and Uttley, A. M. (1953) Pattern discrimina&n and the visual cortex. Nature, 2 71, 387388. Simon, H. A. (1980) How to win at tweuty qucvtions with Nature. In R.. A. Cole (ed.), Perceptionand Reduction af Fluent Speech, Hillsdale, NJ, Erlbaum.
Cognition, 10 11981) 17-23 @Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
17
The ~ncept of working memory: A view of its current state and probable future development ALAN BADDELEY* MRC Applied Psychology Unit
In the early 197Os, Graham Hitch and I began a series of experiments which aimed to answer the apparently simple question ‘What functioJl does shortterm memory serve?’ (Baddeley and Hitch, 1974) We assumed, as had many others before: us that STM serves as a working memory, a temporary storage system that plays a crucial role in many information processing tasks ranging from speech comprehension to arithmetic and from learning to complex reasoning. It followed that if we were to load up STM with some secondzy task, then there would be little capacity left for comprehending or calculating. ‘learning or reasoning and hence the subject’s performance on these tasks would be dramatically disrupted. WC tested this view, using immediate memory for digit strings as our secondary’ task and requiring our subjects to remember strings of six digits at the same time as performing tasks involving verbal reasoning, comprehending prose or learning lists of words. We did obtain performance decrements, but we were impressed by how little performance was disrupted by a near-span dkgit load that should have almost tot&& occupied STM. In attempting to understand our results, we began to modify our view of STM and to replace it with the concept of working memory (WM). In doing so, we moved away from a strategy of exploring the capacities and implications of a single hypothetical structure (STM) and towards a more functional analysis, Working Memory refers to the role of temporary storage in information processing, The concept of working memory The concept of WM with its implication of a limited capacity system for holding and manipulating information has motivated research in a number of areas. In my own case these have included the study of reading (Baddeley, 1979: ---_*This note was written at the University of Guelph; the support of their short-term visiting professorship scheme is gratefully acknowledged. I wish to thank Tim Shallice for his comments on the manuscript. Reprint requests should be sent to: A. Baddeley, MRC Applied Psychology Unit, 15, Chaucer Road, Cambridge, England.
18
Ah
Buddeky
Baddeley and Lewis, in press; Baddeley, Eldridge and Lewis, in press), while Daneman and Carpenter (1980) have shown that a measure of WM capacity provides an excellent predictor of reading ability in college students, a result which we have subseciuently replicated over a wider range of age and ability. An as-yet unpublished series of experiments on the role of WM in retrieval from LTM suggests that whereas WM is important for learning, the process of searching and retrieving from LTM may be largely automatic. Hitch has used the concept highly successfully in studying the processes underlying simple arithmetic (Hitch, 1978), while Broadbent (in press) employs the concept in exp!oring the role of organization in memory. Finally, an impressive and ingenious series of experiments by Rabbitt, Cohen and their co-workers have shown that much of the cognitive deterioration that accompanies normal aging can be attribu,ted to a decrement in the performance of working memory (Rabbit& in press), There is no doubt then that the concept of WMis proving fruitful. However, there is a danger that its very success may leati to its abuse. Given almost any poorly understood performance decrement, it is possible to attribute it to the inadequate performance of WM. Hence, if the concept is to continue to be useful, it is important that an attempt is made to ensure that it is not simply used as a label for one’s ignorance of the underlying casse of a given decrement. Perhaps the best way of ensuring that it does not become a vacuous cat&-all concept is to develop and test specific models cf WM which will al1o.wtestable and non-trivial predictions. A model of working memory In order to account for our various results, Graham Hitch and I formulated a simple model which we hoped would provide a frE;nework for a more deta.iled analysis of WM. The model subdivided WM into three components. The Cenrral Executive which formed the control centric of the system was assumed te select and cperate various control processes. It was assumed to have a limited amount of processing capacity, some of which could be devoted to the short-term storage of information. It was able to offload some of the storage demands on to subsidiary slave systems of which two were initially specified, namely the Articulatory Loop which was able to.maintain verbal material by subvocal rehearsal, and the Visuo-SpatialScratch Pad which performed a similar function through the visualization of spiatial material. The system we proposed was not a predictive model; we were sure that WM would prove to be far more complex than our original conception, and that &en the state of our knowledge, any attempt to make a rigidly specified
The concept of working memop
19
predictive model was bound to fail. What we proposed was much more in the spirit of a tentative map of new terrain, giving broad guidelines and suggesting areas for more detailed exploration. The evaluation of this type of theory rests on its fruitfulness in generating new knowledge and fresh insigilts. How successful has our model been? The articulatory
loop
This is the most extensively explored component of the WM model. In its sriginal form, the articulatory loop was assumed to resemble a tape lop of limited dllration which could be used to store any information that could be articulated. !t has the advantage of neatly tying together results indicating that immediate memory span is susceptible to the effects of both phonological similarity and word length and that both these effect3 vanish when subvocal rehearsal is suppressed by requ&ing the subject to articulate repeatedly some irrelevant word such as ‘the’ or ‘double’ (Baddeley and Hitch, 1974; Baddeley, Thomson and Buchanan, 1975). The concept of an articulatory loop has proved useful in investigating other areas. For example, Ellis and Hennelly (1980) used the concept to expiain wbJf rht: norms for the digit span in the Welsh language version of the WJSC WH ( consistently lower than the English language equivalent. Welsh digits take !onger to articulate than English, and when this factor is eliminated by articulatory suppression, or allowed for by using a time-based measure of span, the Welsh-English difference disappears. As predicted by Hitch’s work on WM and arith.metic, the longer spoken duration of Wels;. ligits also leads to an increase in certain errors in mental arithmetic. Nicholson (in press) explored the hypothesis that the increase in children’s memory span with age might stem from the development of the articulatory loop. He showed a very close relationship between age, memory span and the: rate at which children were able to articulate digits. Research on the role of the articulatory loop in fluent reading has shown that it is not essential for comprehension (Baddeley, 1979), but provides a supplementary source of information that may be important when a pigh level of accuracy is required (Baddeley, Eldridge and Lewis, in press; Levy, 1978). A number of results suggest that the articulatory loop may play an important role in learning to read (Conrad, 197%; Liberman, Shankweiler, Liberman, Fowler and Fischer, 1977) although the evidence is still not conclusive. Taken overall then, the concept of an articulatory loop does appear to be proving a fruitful one. Attempts to analyse the articulatory loop in more detail however, have thrown up a number of apparent anomalies. For example, although articulatory
261
AlanBaddeley
suppression removes the effects of phonemic similarity and word lap memory span for visually presented items, with auditory preantati phonemic similarity effect reappears despite suppression, indicati nemic coding is not dependent on the process of articulation. It thought that this was also true of the word length effect, but a recent study has showr, that provided articulation is prevented throughout both presentation and recall no word-length effect occurs, even with auditory pr tation (Baddeley, Lewis and Vallar, unpublished). It appears then that wordlength effect reflects the process of sub-vocal rehearsal--1on poor recall because of the slower rate at which the memory freshed by rehearsal. The nature of the phonemic similarity effect remains unclear. It may reflect the speech-based bystem which holds the codes that feed the articulatory loop. If so, we must conclude that such codes can be set up either auuitorily or by articulation, but not purely visually. Alternatively, the phonemic similarity effect may result from either of two separate cocies, one articulatory and the other auditory. The similarity betw articulatory and auditory features makes it difficult in practice to disti ish between two such codes. Further evidence for the existence of a non-articulatory phonological co& comes from studies of reading (Baddeley and Lewis, in press; Besner, Davies and Daniels, in press). Articulatory suppression was found to have little effect on either the speed or accuracy with which subjects could make rhyme judg ments. Try reading the next sentence while repeating the word ‘double’ to yourself; most people can still ‘hear’ an ‘inner voice speaking the words they read. It seems likely that this non-articulatory phonological code will be actively investigated in the next few years; the discovery of a technique for sup pressing this code would be very helpful qt this point, and might indicate whether it should be considered as an auditory image, or as an abstract pm articulatory code that is sufficiently deep to be unaffected by concurrent articulation. The Visuo-Spatial Scratch Pad
Although the experiments supporting the concept of a visuo-spatiai scratch pad (VSSP) were carried out at the same time as the initial articulatory loop studies they have only just appeared in print (Baddeley and Liberman, 1980), and have so far generated little further research. The scratch pad concept is however, consistent with the flood of research on spatial imagery that has appeared in recent years. It further suggests that the dichotomy between prop ositional an& analogical views of imagery may be a false one. It is for example, quite likely that the scratch pad is a device which takes propositional codes
The concept ofworking memory
21
memory and manipulates and displays them via an analogical rovides an appropriate framework for integrating ~opu~a~2~dby Shepard and his co-workers with work on individual diffeqences in imnce tests (Hunt, 198Ot. The major existence of a second imagery system ~~n~e~led with eon-spatial, pictorial or pattern information. It is clear ly that such non-spatial factors as colour, texture and shape Phe can is also clear experimentally that a concurrent spatial task will ry, but does not diminish the powerful effects of ima for words (Baddeley and Liberman, 1980). While evidence for the disruption of such a pictorial system has been reported (Atwood, t 97 1; Janssen, 1976) effects tend to be small and difficult to replicate. A good technique fordisruptingvisual but not spatial imagery would represent a major breakthrough in this area.
This represents the most important but least understood of the three initial components of WM, and presents the most difficult problems both technically and conceptually.An adequate theory oftheCentra1 Executive wour3 probably include not only a specification of its method of manipulating cor)trol r>rocesses and integrating the growing number of peripheral systems, but would also require an understanding of oelectfve attention and probably of the role and function of consciousness. In order to make someprogress, we have adopted a policy of beginning with the more ripheral components of WM,gradually attempting to separate executive, The hope here is that the more out further su off as tractable subproblems, the greater components that can be separa the chance of reducing the f Executive to a problem of manageable propsrtions. Howmx, it is at present probably fair to regard the Central Executive as the area of our residual i norance about WM: it is at present unfortunately very large and very important. However, some progress has been made. We originally assumed that a concurrent digit load impaired performance on tasks such as comprehension and reasoning because any digits that exceeded the capacity of the articulatory loop had to be held in the central executive; each such digit was assumed to demand some of the limited processing bapacity available. Two unpublished subsequent results argue against this. First, a parametric study of the effects of dig3 load on reasoning showed thrt
22
AkanBaddeley
some subjects were able to hold seven or eight digits z4G!e performing a complex verbal reasoning task with no apparent decrement in speed or accuracy. Although all subjects showed a breakdown in performance if the load were large enough, they seemed to be able to handle many more digits than we had assumed could be maintained in the articulator/ loop. This pattern of breakdown suggests a separate verbal memory component that does not interfere with reasoning until its capacity is exceeded. Such a view is consistent with the neuropsychological evidence reviewed by Shallice ( 1979). A similar conclusion is suggested by a detailed examination of the interao tion of the memory and reasoning tasks. Suppose we assume a pool of general processing capacity that must be shared between the two tasks. For a given level of difficulty, any tendency to assign more capacity to reasoning will aid that task but only at the expense of the memory task, and vice versa. In short, performance on the two tasks should be negatively correlated. Examination of the data suggests exactly the opposite; fast reasoning goes with good digit recall, and vice versa. Such a pattern of results is inconsistent with the assumption of a pool of general processing capacity. It suggests rather two separate though related systems, a memory system and a controller which runs it. Provided the memory system is coping, it will place few demands on the controller, which can therefore use its available capacity for performing the reasoning task. Once the memtjry system becomes overloaded however, it requires the controller to bring in further resources in an attempt to avoid breakdown and consequent loss of the stored mat+al. The more attention that is devoted to supporting the memory component, the less is available for reasoning, and the slower and more erro: prone is the reasoning performance. It seems probable then that we should distinguish between a verbal memory component and an attentional component. Should this memory component contLrue to be regarded as part of the Central Executive? Probably not, in which case it could be argued that the Central Executive is bec.oming increasingly like a pure attentional system. Whether or not this proves to be so, it seems likely that any adequate model of WM will also have to be a model of attention.
References Atwood, G. E. (1971) An experimental study of visual imagination and memory. Cog. Pgwhol.,2,290-
29!L
Baddeley,A. D. (1979) Working memory and reading. In P. A. Kolers, M. E. Wrolstad and H. Bouma (eds.h Recessing of K!si& hnguwe. New York, Plenum. Baddeley, A. D., Eldridge, M., and Lewis,V. J. (In press) The role of subvocalization in reading. Q. J.
expez &vctwi,
r”he concept of worting memory
23
Baddeley, A. D., and Hitch, G. J. (1974) Working memory. In G. A. Bower (ed.), ThePsychology of Learning and Motivation,Vol. 8, New York, Academic Press. Pp. 47-90. Baddeley, A. D., and Lewis, V. J. (In press) Inner active processes in reading: The inner vo’ce, the inner ear and the inner eye. In A.M. Lesgold and C. A. Perfetti (eds.), InteractiveProcessesin f?eading. Hillsdale, NJ., Erlbaum. Baddeley, A. I)., and Lieberman, K. (1980) Spatial working memory. In R. Nickerson (ed.),.Ittention and Performance VIII, Hillsdale, NJ., Erlbaum. Badddey, A. D., Thomson, N., and Buchanan, M. (1975) Word length and the structure of short-term memory. J. verb. Learn. verb. Behav., 14, 575-589. Besner. D., Daviers, J., and Daniels, S. (In press) Phonological processes in reading: The effects of concurrent articulation: Q. J. exper. Psychol. Broadbelt, D. E. (In press) From the percept to the cognitive structure. In J. B. Long and A. D. Badde’iey (eds.), Attention and PerfornzanceZX, Hi&dale, NJ., Erlbaum. Conrad, R. (1972) Speech and reading. In J. F. Kavanagh and 1. G. Mattingley (eds.), Languageby Ear and by Eye. Cambridge, Mass., MIT Press. Daneman, M., and Carpenter, P. A., (1980) Individual differences in working inemory and reading. J. verb. Learn. verb. Behav., 19, 450-466. Ellis, N. C., and Hennelley, R. A. (1980) A bilingual word-length effect: Implications for intelligence testing and the relative ease of mental calculation in Welsh and English. Brit. J. Psychol., 7:, 43-52. Ernest, C. H. (1977) Imagery ability and cognition: A critical review.J. ofMent. Imag., 2,181-X16. Hitch, G. J. (1978) The role of short-term working memory in mental arithmetic. Cog. Psychol., 10, 302-323. Hunt, E. (1980) Intelligence as an information-processing concept. kit. j: Psychol., 71, 449-474. Janssen, W. H. (1976) Selective interference during the retrieval of visual images. Q. J. exper. Psychol., 28. 535-539. Levy, B. A. (1978) Speech analysis during sentence processing: Reading and listening. Visiblelanguage, 12, 81-101. Liberman, I. Y., Shankweiler, D., Liberman, A. M., Fowler, C. and Fischer, F. W. (1977) Phonetic segmentation and recoding in the beginning reader. In A. S. Reber and D. Scarborough (eds.), ToHnrrrdsa Psychologyof Reading. Hillsdale, NJ.., Erlbaum. Nicholson, R. (In press) The relationship between memory span and processing speed. In. M. Friedman, J. P. Das and N. O’Connor (eds.), Intelligenceand Learning, New York, PRenum Press. Rabbitt, P. (In press) Cognitive psychology needs models for changes in performance with age. In J. B. Long and A. D. Baddeley (eds.),. Attentioll and Performance IX, Hillsdale, NJ., Erlbaum. Shallice, T. (1979) Neuropsychological research and the fractionation of memory systems. In LG. Nilsson (ed.), Perspectivesols Memory Reseurch. IHtisdale, NJ., Erlbaum. Shepard, R. N., and Met&, J. (1971) Mental rotation of three-dimer,sional objects. Sci., 271, 701703.
&ition, 10 (1981) 25-32 @ Elsetier Sequoia S.A., Lausaone - Printed in The Netherlands
25
Phonetic features and acoustic invariance in speech SHEILA E. IBLUMSTEIN* Bmwn University KENNETH
N. STEVENS
Massachusetts Institute of Technology
Most researchers i:n speech perception have assumed that a p:rimary perceptual unit is the phonetic segment or the distinctive feature (Liberman, Cooper, Shankweiler, and Studdert-Kenneoy, 1967; Stevens and House, 1972; Studdert=Kennedy, 1976; Pisoni, 1978). In this view, a word can be decomposed into individual segments, and the segments, in turn, can be further analyzed into component se&& of distinctive features (Jakobson, Fant and Halle, 1963). The features, in effect, organize the segments into natural classes, each characterized by a particular feature. In fact, traditional linguistic analyses have assumed that the sound system of natural language is structured in terms of individual segments and features. Typically, a natural class is defined as a set of speech sounds which share a particular feature, and in which fewer features are needed to characterize the particular class than any one member of that class (Halte, i964). The empha.:.ts tin the segment and feature rather than the syllable or word is based on evidence from linguistic theory, synchronic and diachronic linguistics, and linguistic performance. With regard to linguistic theory, the use of -tie segment and feature provides a principled means of accounting for significant linguistic generalizations. For example, a language is more likely to have phonological changes described in (a) than in (b). In (a) [p ] becomes [b 1, [t] becomes [d] , and CkJ becomes [gl when they occur between any two vowels, (b) [p] becomes [ I], [t] becomes [z) , and [kl becomes [WI when they follow the syllables [ba- I%,[ ti-] , or [ lu-1 \a) the changes can be characterized by a phonological generality, i.e., voiceless stops become voiced intervocalically, and in fact changes of this typeoccur commonly in natural language. In contrast, the changes described in (b) are virtually non-existent in natural language, are phonologically ad hoc, and can not be described in a systematic or general way. lflrfswork was supported in partby GranEsNS 15123 and NS 04332 from the National Institutes of Health. Repxint requests should be sent to S. E. Bhunstein, Department of Linguistics, Brown Univex&y, Providmce,R. 1.02912, U. S. A.
26
SheilaE. Bkmsteirsand KennethN. Stevens
Synchronically, natural language rules typically involve segments and features, and not syllables or lexical items. For example, the various forms of the English plural are determined by the voicing and manner characteristics of the preceding consonant: the plural is realized as [P~z]in the context of a sibilant, [s] in the context of avoicelessconsonant, and [z] in the context of a voiced consonant. Similarly, diachronic or hisitorical linguistic changes usually affect segments or natural classes of speech sounds. Grimm”s law is a prime example of suchachange. In this case, the correspondance betwleen proto-Indo-European (PIE) and Germanic (Gmc) can be ch’aracterized in terms of feature changes in particular sound classes. Thus, PIE unvoiced stops are realized in Gmc as unvoiced @rants, voiced stops as unvoiced stops, and voiced aspirates as voiced fricatives or spirants. Finally, psycholinguistic evidence clearly supports the importance of the segment in speech. For example, speech production errors in aphasia are primarily segmental in nature (Blumstein, 1973) and slips of the tongue commonly affect segments a:ldi features (Fromkin, 1973). Thus, evidence from the sound structure of natural language suggests that the segment and/or feature is an independent component or unit in the representation of utterances in natural language. Acoustic invariance for phonetic features and sources of variability in segments. Acoustic analysis of different phonetic segments that have a feature in common shows that the detailed acoustic manifestation of the feature can be affected by the phonetic context in which the feature occurs, the vocal tract size of the speaker, and the speaking rate. Thus, fc’aexample, it is possible to find certain acoustic parameters in the sound in the vicinity of the release of the alveolar stop consonant [d] that depend upon the following vowel. On the basis of such observations it might be Iconcluded that the acoustic attributes identifying the consonant as an alveolar are entirely context-dependent, (Liberman et& 1967).Alternatively,itmay be possible to find, for each feat.ure!, a unique acoustic property that is not sensitive fo these context-dependent Variations, i.e., a property that ignores these variations and focuses on the invariant attributes of the segments that contain this feature. In order to clarify this concept of invariance, it is necessary to state what we mean by the term property. We postulate some kind of process or model that gives a maximum response magnitude when the parameters of the sound have particular ranges of values, and for which the response drops sharply when the parameters fail outside these ranges. The sound is said to have the given property if this process gives a substantial response to the sound. The
Phoneticfeaturesand acousticinvariancein speech
27
implication is that the auditory system is equipped with property-detecting -mechanisms of this kind. A property, then, corresponds t,-, a natural and simple acoustic attribute having a particular frequency-time structure to which the auditory system can respond selectively and distinctively. What is the nature of these properties, and where are they observed in the speech stream? We view the, properties as being integrated, in the sense that they encompass characteristics that spread over a range of frequencies and attributes that span a window extending over some tens of milliseconds of time. For example, a property tends not to be defined in terms of the absolute frequencies of particular spectral peaks but rather in terms of the gross shapes of spectra sampled at par-titular points in time or over particular time regions (Stevens, 1975; Blumstein and Stevens, 1979; Stevens and Blumstein, 198 1). The search for an invariant acoustic property corresponding to a phoneti:: feature require-. a determination of where to measure the property, and what proprrty to F :asure. This search can be guided in part by known characteristics of the response of the auditory system to speechlike sounds (as discussed later) and in part by ‘considering how lack of invariance could occur in the signal and therefore how one might select properties at points in time that are expected to be insensitive to these variations. Lack of invariance appears to arise for at least three reasons. First, there are points in time where the articulatory structures are in transition from one target configuration or state to another. The acoustic parameters at these points are inevitably dependent on both the preceding and the following target configuration, and consequently are determined by features for two or more adjacent segments, Secondly, even at a point in time where certain articulatory structures are close to the target configuration or state defined by a particular feature, there may be other aspects of articulation that are not constrained by the features that define the contrasts for the language. Thus, for example, nasality is not a feature that defines contrasting vowel classes in English, and, consequently, when a vowel is followed by a nasal consonant, a speaker is free to anticipate the consonant nasalization by nasalizing the preceding vowel. A. third reason for lack of invariance is that different speakers have different vocal tract lengths and shapes, and consequently there may be variation in the natural frequencies of the vocal tract from one speaker to another for a given speech sound. Sources of variability such as these need to be understood and taken into account as one proceeds to search for invariant acoustic correlates of the features. Implications of a theory of acoustic invariant The major claim of a theory of acoustic invariance is that invariant acoustic pro$ertias can be derived directly from the acoustic signal, and these properties
28
Sheila E. Blumstein and Kenneth N. Stevens
correspond to the phonetic dimensions which ultimately form the inventory of speech ‘sounds used in natural language. Such a view provides an explicit characterization of the universal set of features, and, in particular, of the phonetic dimensions delineating natural classes. Ifit is the case that invariant properties structure phonetic dimensions, then such invariance provides an instantiation of particular feature systems. A theory of acoustic invariance at the very least provides a listener with a mearr:, for organizing the sounds of his language into classes. Orce these clas ses are established, they are available for representing lexical items and for serving as a b+s for describing phonological regularities or constraints that exist in the language. A stronger claim, however, is that, during ongoing speech perception, we use these properties, on-line, to identify features in the utterances, and, as a consequence, to identify lexical items that are represented in memory as matrices of features. If such a hypothesis concerning on-line use of invariant properties is correct, then it would suggest that complex computational procel’ures to extract these dimensions, as delineated in models of analysis-by-synthesis (Halle and Stevens, 197?), and the motor theory (Liberman, et al., 1967) would not be required. In fact, there would be a fairly direct relation between the acoustic signal itself and the phonetic dimensions of speech. Further, such a view could account for the ability of the infant to perceive the phonetic 2imensions of speech with only a limited exposure to language in a manner similar to the adult (Eimas, Siqueland, Jusczyk, and Vigorito, 197 1). If the infant is born with ‘iheability to extract appropriate acoustic properties in sound stimuli in general, then the phonetic system is in effect structured for him. Nevertheless, it is. probably the case that in ongoing speech processing the perceptual system makes use of a number of acoustic attributes, including onset frequencies of formants, transitions of formants, and vowel formant frequencies, as well as the invariant properties (cf., Liberman et al, 1967). Whether these, invariant properties are in themselves perceptually primary, in the sense that they are the most salient acoustic cues used by the listener and thus ‘direct’ phonetic categorization, has yet to be determined. However, even if these properties are only among several acoustic dimensions contributing to phonetic categorization, the hypothesis is that they still provide the organ nizational framework for the phonetic system of natural language, as well as providinb the perceptual system with a direct representation of the structural prcperties inherent in natural language (Jakobson, et al., 1963; Stevens and BP‘mstein, t 98 1). Evidence for a theory of acoustic invariance Evidence for a theory of acoustic invariance must be based in part on acoustic
an?;ysis of the speech wave for different speech sounds in a variety of pho-
Phonetic features and ucou~ticinvariance in speech
neticcontexts.
isto
29
an invariant property can be defmed for each feature, and to oetermine whether the same property can be mesured in a number of different speech tokens in which this feature plays a role. Evidence concerning the role played by these invariant acoustic propeeeh perceptitBn process must come from speech perwhich listener responses are obtained for speech and speechlike stimuli ntain these acoustic properties. test ofa theory ofinvariance is to examine the acoustic properties and the ~r~epti~n for features that specify place of articulatie;r nsonants. it is for this phonetic dimension that experimental data that, for a given place of articulation, there are some acoustic garameters ly dependent upon the fohowi Furthermore, there is t the perception of the phtlne sion of place of ar!iculatisn can, under certain circumstarrees, make use of context-dependent cues (Liberman et al., 1967). For example, both acoustic measurements and studies of speech perception have shown that, for a given place of articulation, there are acoustic parameters such 35 direction of formant transitions, frequency of burst, or onset frequencies of L’ormantj,that are contextdependent ( Yooper, Calattre, Liberman, Borst, and Cerstman, 1952; Schatz, 1954; Delattre, Liberman, and Cooper, 1955). New data have recently emerged, however, showing that, if the acoustic property is appropriately selected, acoustic invariance may be specified for a given place of articulation independent of speaker, vowel context, syllable position, and voicing characteristics (Stevens and Blumstein, 1978; Blumstein and Stevens, 1979). The acoustic property that provides this result is obtained f 20-odd milliseconds from the spectrum sampled over a brieftime interv at the release of the consonant. At this point in time t articulators are close to being m the target positions appropriate for the stop consonant, and thus the acoustic propertie &eating place of articulation are least likely to be affected by adjacent ents. The property can be characterized in terms e for each of the labial, alveolar, and velar of a unique, press spectrum of velars there is a prominent spectral peak places of articulation. In the in the midfrequency region, whereas for labials and alveolars there is a diffuse spread of spectral energy over a range of frequencies. This diffuse spee trum shows substantial energy at high frequencies for alveolar’s but not for labials. Attempts have been made to quantify these perties in terms of a set of spectral templates (one for each place of art measured spectra for individual tokenscan be matched (Blumstein and Stevens, 1979). Data from a thousand-odd examples of consonant-vowel syllables produced by several different speakers showed that about 85 percent of the tokens for each place of articulation matched the corrc :t template independent of the vowel context or the voicing characteristics of +%econsonant. e tirsk
deWmine
)
whether
30
SheilaE. Bhmsterin andKennethN. Stevens
A series,& perceptual experiments was carried out to determine whether listeners use these gross spectral properties, sampled at the consonantal release, in identifying place: of articulation for stop consonants. These experiments showed that (1) listeners could identify the consonantal place of articulation when only the initial lo-20 msec of the syllable was presented (indicating ,that information about place of articulatfon is carried in this brief time interval at consonant release) (Blumstein and Stevens, 1980); and (2) manipulat‘qn of the spectral characteristics at the onset of a synthetic syllable caused listeners to shift their identification of place of articulatron in ways that were predictable from measurements of the gross spectrum shape using the templates described above (Stevens and Blumstein, 1978; Blumstein and Stevens, 1980). While the features that specify place of articulation for stop consonants have been examined in some detail, as we have just noted, there is substantial evidence of a more qualitative nature to indicate that there are invariant acoustic correlates for a number of other features (Jakobson, et al., 1963; Stevens and Blumstein, 198 1; Stevens, 1975, 1980). For example, invatiant acoustic properties have been postulated for the following features: contin:;&tf (characterizing the class of fric$ives in contrast to other classes of segments): strident (distinguishing [sH&j 1 from other consonants in English), nasal, voicing, and others (Stevens and Blumstein, 198 1). There is also evidence for invariant acoustic properties that define tongue position for vowels (Stevens, 1980). For many of these properties there is perceptual evidence to indicate that the speech perceptior system gives a distinctive response when the ap propriate property is present in the sound (for discussion, see Miller, Wier, Pastore, Kelly and Dool’ing,1976; Pisoni, 1977 ; Stevens and Blumstein, 198 1). Further systematic research is needed, however, in order to specify these properties in detail. This additional research needs to (1) examine the acoustic properties corresponding to the features as they occur in various phonetic environments and in d.ifferent languages, and (2) ascertain, through perceptual experiments, the role: played by these propertiesin identifying the features for the listener. The study of the acoustic properties associated with particular features in different languages is of especial interest because it is then possible to observe the acoustic manifestation of a given feature as it appears in combination with a variety of different patterns of features. For example, in English or m French there is only one place of articulation for stop consonants produced with the tongue blade -alveolar for English and dental for French. On the other hand, there are three coronal stops in ?~aIayalam--dental, alveolar, and retroflex. Is there an acoustic property that all of these coronal consonants have in common? What are the properties that distinguish between different coronal con-
Phmetic features d amustisinvariance
it? speech
31
uestions such as these can only be answered through cross-language rrent research s ests, for example, that the diffuse-rising spectral property at onset derive studies of alveolar stops in English may netid some revision if it is to e a general characteristic of coronal consonants in different Ian Further rese is also needed in order to elaborate on the general nature of the property-detecti mechanisms that are postulated to be part of the speech processin Do these mechanisms operate for auditory signals in general, and are the sounds of speech so structured as to take advantage of these property-detecting capabilities? There is evidence from psychoacoustic experiments that these property-detecting mechanisms operate to facilitate the organization of sounds into classes (see, for example, discussion in Chistovirh, 1971 and Stevens, in press), but more data of this kind are needed in order to explore systematically the response of the auditory system to sounds wihh speechlike properties. Can data from auditory physiology shed light on the processing of speechlike sounds by the auditory system (Chistovich, 197 I) and provide evidence for a predisposition of the auditory system to respond selectively or in qualitatively different ways to sounds that have contrasting properties? Work in this area, using animal preparations in which the peripheral auditory system is similar to that of the human, is anly beginning to provide data on how the peripheral auditory system responds to sounds with speechlike properties (Delgutte, 1980; Kiang, 1980; Sachs and Young, 1980). At the very least, this work will lead to a specification of the constraints imposed by the peripheral auditory system on the representation of auditory signals at a higher level.
Blumatein, S. E.(1973) A Phonologicul Inverfigufion ofAphasic Speech. The Hague, Mouton. Blumsteln, S. E.and Stevens, K. N. (1979) Acoustic invariance in speech production: evidcncc from
measurements of the spectralcharactcristics of stop wnsonants.J. urwsf. Sot. Amer., 66, 1001’1017. Blumstoin, S. E., and Stcvans, K. N. (1980) Petccptual invatiancc and onset spectra for stop consonants in dlffc *ent vowel environments. J. acousf. Sot. Amer., 67,648-662. Cb.stovich, r A. (1971) Auditory plrocessing of speech stimuli: evidence from psychoacoustics and neurophysiology. In Bweedings of VIZ International Congresson Acoustics. Vol. I. Budapest, Akaddmiai Kiad6, Pp. 27-42. Cooper, F. S., De&We, P. C., Liberman..,A. M., Borst, -.1 M., and Gerstman, L. J. (1952) Some experiments on the perception of synthetic speech sounds. J. acousf. Sot. Amer., 24, 597-606. Delattre, P. C., Liberman, A. M., and Cooper, F. S. (1955) Acoustic loci and transitional cues for con. sonants. J. acousf. Sot. Amer., 2?,769-773.
32
Sheila 5+.Bkmstein
and Kenneth N. Stevens
Delgutte, B. il980) Representation of speech-like sounds In the discharge patterns of auditory=nerve fibers. J, acoust. 5%. Amer., 68,843-857. Eimas, P. D.,Siqueland, E. R., Jusczyk, P. and Vigorito, J. (1971) Speech perception in infants. SCi., I ;‘i ,303-306. Fromkin, V. A. (1973) (ed.) Speech Errors as LinguisticEvidence. The Hague, Mouton. Halle, M. (1964) On the bases of phonology. In J. A. Fodor and J. J. Katz (eds.), The Structure of Lun. guage. Englewood Cliffs, NJ, Prentice-Hall. I‘p. 324-333. Halle, M. and Stevens, K. N. (1972) Speech recognition: a model and a program for research. In J. A. Fodor and J. J. Katz (eds.), The Structure of Languuge, Englewood Cliffs, NJ, Prentice-Hall. Pp. 604-612. Jakobson, R., Fant, G., and HaUe, M. (1963) tieliminuries to Speech Analysis. Cambridge, Ma&, MIT Press. Kiang, N.Y.-S. (1981)) Processing of speech by ihe auditory nervous system. J. acoust. Sot. Amer., 68, 830-835. Liberman, A. M., Cooper, F. S., Shankweiler, D. P. and Studdert-Kennedy, hi. (1967) Perception of the speech code. Psychoi. Rev., 74,431-461. Miller, J. D., Wier, C. C., Pastore, F.. E., Kelly, W. J., and Dooling, R. J. 11976) Discrimination and labelling of noise-buzz sequences with noise-lead times: an example of categorical perception. J. acoust. Sot. Amer., 66,1001-1017. Pisoni, D. B., (1977) Identification and discrimination of the relativesnset time of two component tones: implications for voicing perception in stops. J. acoust. Sot. Amer., 61, 1352-1361. Pisoni, D. (1978) Speech perception. In W. K. Estes (ed.) Handbook of Learning and Cognitiucf% cessea Vol. 5. Hillsdale, NJ., Erlbaum. Pp. 167-233. Sachs, M. B. and Young, E. D. (1980) Effects of nonlinearities on speech encoding in the auditory nerve.Z Rcoust.Sot. Amer., 68, 858-875. Schatz, C. D. (1954) The role of cont=xt in the perception of stops. Lang., 30.47-56. Stevens, K. N. (1975) The potential role of property detectors in the perception of consonants. In G. Fant and M. A. A. Tatham (eds.), Auditory Analysis and Perceptdon of Speech. London, Academic Press. Pp. 303 -330, Stevens, K. N. (1980) Acoustic correlates of some phonetic categories. J. ucoust. Sot, Amer., 68,836842.
Stovens, K. N. (in press) Constraints imposed by the auditory system on the properties used to classify speech sounds: data from phonology, acoustics, and psychoacoustics. In T. F. Mycrsand J. Lavtx teds.), Proceedings of the International Symposium on rhe Cog&Sue RepresentarM of Speech. New York, Elsevier North-Holland. Stevens, K. N. and Blumstein, S. E. (1978) lnvariant cues for place of articulation in stop consonants. J. acoust. Sot. Amer., 64, 1358-1368. Stevens, K. N. and Blumstein, S. E. (1981) The search for invarrsnt acoustic correlates of phonetic fWures. In P. D. Eimas and J. L. Miller (eds.), PerspecCiveson the Study of Speech, Hilldale. NJ., Erlbaum Associates. Pp. l-38. Stevens, K. N. and House, A. S. (1972) Speech perceptiorr.In J. V. Tobias (ed.), FounduHons of Modern Auditory Theory. Vol. 2. New York, Academic Press. Pp. l-62, Studdert-Kennedy, M. (k976) Speech perception. In N. J. lass (ed.), Contempowy Issues in &per& mentul Honetics. New York, Academic Press. Pp. 243-294. Zhukov, S. Ya., Zhukova, M. G. and Chistovich, L. A. (1974) Some new concepts in the auditory analysis of acoustic flow. Sov. Phys.Acoust., 20,237-240.
@Worz, IO (1981) 33-38 @ Blsevier Sequoia S.A., Lausanne - Printed :mThe Netherlands
33
Chomsky without language ALBERT S. BREGMAN+ iUcGil! University
After reading Chomsky’s Aspects oj’ the 77~0~ of Syntax (1965), I became intrigued by the idea of composition. Consider the sentence “The stone was thrown by the angelic little boy”, In the Aspects approach, the passive voic, was viewed as being a simple part of the sentence at the deep structural level, but as corn@ out on the surface structure as a word-order arrangement, Many features of sentences were treated by Chomsky in this way. Furthermore, the exact form that a particular deep pattern would take when it appeared in the surface structure of the sentence would differ, depending on the other deep patterns which appeared in the same sentence; yet at the deep structural level it was always the same. T-his made me think of a problem posed by William James which can be stated as follows. If a piece of behavior or thought is constructed out of a set of distinct units, as the associationists had claimed, how can two observations be explained: (a) the “same” act or thought never reoccurs in exactly the same form on two occasions; (b) the so-called units flow into one another and overlap; we cannot determine the exact end of one an;! the beginning of another. It seemed that Chomsky had addressed these two problems as they appeared in syntactic structures, and had solved them by postulating a double level of description. His deep structural description Pabelled the separate identity of each syntactic form and made this identity available to rules of composition. His surface description concerned the patterning that resulted from the composition, 1 patterning in which the deep forms were modified, merged and modulated. Thus it seemed that Chomsky could eat his cake and have it: he could have a rest&ted number of invariant forms at the deep level and an infinity of unique patterns at the surface. This situation was made possible by ,the fact that. the deep forms were put together not by the simple concatenafion proposed by the associationists but by a rather t,omplicated process of composition. (hamsky believed that thisspecial aspect of language, deep structure, made it impossible to learn language by association, a process which simply concatenated elements. Therefore the obvious capacity of humans to both perceive *Reprint requests should be sent to Dr. A. S. Bregman, Department of Psychology, McGill University, 1205 Doctew Penfield Avenue, Montwal, Quebec H3A lB1, Canada.
34
A. S. Bregmun
and produce language pointed to the existence in humans of a powerful p:.ocessing mechanism, a language acquisition device, specifically designed to extract the deep regularities of language from observed sentences. This capacity, in the opinion of Chomsky, marked language as unique and a special sign of our humanity. The idea that human psychological events are created by the composition of underlying abstract elements was wonderfully attractive and the more I thought of it, the more it seemed to apply to many aspects of human psychology outside the field of language. Perception, it seemed, could also be seen as a process of composition. In this view, our perceptual process composes some basic notions to generate a structure that matches the current input. An example, drawn from vision, goes as follows. When we perceive a shiny table top, the distribution of light that hits our retinas is interpreted as arising from at least two distinct (deep) aspects of the fable top, its surface color and its glossiness. Furthermore, these two ideals must be taken in composition, not one at a time, to account for the pattern of color that hits our eyes, how it changes as our head moves, and how it is slightly different in the two eyes. The resulting percept can be seen as having two levels, a deep one in which the color and &ssinet;s are registered as separate featlures of the table top and a surface level in which the actual distribution of light is accounted for. In Chomsky’s system, $ransformational rules were given the job of describing the modifications of the deep patterns as they beca.me realized in a sentence. By anology, the perceptual system can be said to have rules which know (in some sense) how the color and glossiness of surfaces interact in shaping the final distribution of light. We can call these the. rules of composition for deep visual regularities, and can assign the name ideals to the underlying visual concepts themselves. If simple aspects of the world, say surface color and glossiness, compose their effects upon the evidence that reaches our senses, our perceptual systems face the same horrendous detective problem faced by our language system, that of decomposing the evidence back down into its constitutive regularities. The syntactician refers to this as the parsing problem. The problem of finding the deep regularities that have shaped a pattern of data is also the problem faced by the scientists; so our senses must practise scientific explanation in their own modest ways. When they attain two adequate explanations of the data by performing two different compositions of underlying ideals, we have an ambiguous figure. When the perceptual system’s choice of underlying factors is different from the one used by the external world in shaping the perceptual input, then we have an illusion. When internal factors dominate the perceptual compositon so that it fails to be driven by sensory data, we call it imagery, hallucination or dreaming, depending on other aspects of our state of consciousness.
For about twelve years now, I have been trying to apply this concern with composition and its converse, the recovery of regularities, to problems in auditory perception. One of the most fundamental ways in which simple sources of sound can compose their effects upon the wave pattern that reaches our ears is the process of mixing. Outside the psychoacoustic laboratory, what we hear is always a mixture of the effects of different vibrating sources. An adequate perceptual representation of this input will contain, at the deep level, a separate description of each source in terms of such auditory primitive concepts as loudness, pitch, and location, and some not-so-primitive concepts such as the words spoken by one of thevibrating sources. The actual acoustic consequences of the process of mixing will have to be represented in a surface level of description so that it can be tested for adequacy against the sensory input. One consequence of 1:hemixing of two sounds is the masking of one sound by another. The auditory system has a\ way of discounting the effects of masking. Whenever an input can be matched by a perceptual construction in which a loud sound is represented as interrupting a soft sound, the masking of the evidence for the soft sound at that point in time is not taken as implying the discontinuation of the soft sound. During masking, the soft sound is represented at the deep level (and indeed actually heard) as continuing through the interruption. Another consequenceof the mixture of two patterns of sound is that the elements
36
A. S. Bregmm
supply any help in its execution. In fact there is no similarity between 8 pattern of sounds and the muscular responses that produce a similar pattern” (Skinner, 1957, p. 59). While it is possible that an account of the behavior of very young infants may have to be written in a vocabulary that is quite independent from that of perception, the example of humEn language seems to show tha:, at least in this field, the behavior of an adult has to be expressed in the sane concepts as that used to describe percepts. It is important to see, however, that this equivalence is not a special fe’ature of the domain of language. We also perceive meaning in the non-linguistic behavior of others. For exampie, I may see a man’s struggles with a sack Iof potatoes as being organized around the goal of getting the sack onto his back. If I am correctly interpreting his actions, both his behavior and my perlception are being organized by this same goal. In his behavior, the goal organizes the transitions from one movement to the next. For me, my perception of his goal accounts for the transitions between one of his movements and the next and allows me to predict his future moveJz ments. It is possible to argue that any type of behavior is a composition formed from underlying abstract “ideals” and that these very same ideals form the basis for our understanding of the behavior of other people. Just as the ideals in language (the syntactic forms for example) are very abstract, so are those that gener lte iIon-linguistic behavior. Consider the act of grasping. I can grasp with my finger and my thumb, with two fingers, with my knees, with my elbows, with chopsticks, with one knee and a chopstick and so on. The way grasping is instantiatedin my behavior depends on what other elements are appearing at the same time. I must obviously possess a formula that defines grasping, one that is no less abstract or far from the surface of an action than is the passive element in Chomsky’s account of the syntactic form of a sentence. Why are we built this way? To answer this question we need to look more closely at what function the abstract elements or ideals serve. My view is that in any individual situation, a composition of basic idsals is created. in the field of perception, the composition serves as a model of the situation. The model is built from a set of more fundamental hatterns, each 0~1ecapturing some important regularity of our world. Because the model of th:) specific situation is built from a vocabulary that captures fundamental reguiarities, such as surface, slant, object, above, person, goal, :ind so on, the model serves as a powerful calculating system. Each regularity carries with it rules for calculating what may happen next or what actions may be carried out successfully. in the field of action, the composition has a parallel function. It organizes behavilor around abstract action-ideals, such asgrusp, hurry, go to, be polite,
Wmmsky withuut language
37
S& and SOon. Each such action-ideal captures some regularity in our interaction with the physical or social world and represents a type or aspect of coherent amd effective action. One can see, furthermore, that thought sequences are compositions whose structure also derives from ideals. In Dtumanthought, these ideals have enormous variety, ranging from such specific entities as Felix, the cat to such allpurpose organizational patterns as means-end, or classification, and such computational patterns as negativ’eor hygothesizitlg. I mean these itrilicized words to refer to patterns that structure our thought and not to the lords that we use to label such patterns. However, every word of our language links to one of these ideals. The human child’s development of these underlying ideals constitutes what Piaget called “the development of objectivity”, a phrase that he employed to refer to the process by which inner patterns of the mind come to have forms that match the forms of external reality. Any adequate mental system for modelling reality will have to have elements whose own forms and patterns reflect external ones. While the brains of other species and the computational systems which they embody must also have this property, the human brain has developed an astounding capacity to handle a lot of these modelling patterns at once. Furthermore it can extend its intellectual system by adding more ami :rsoreof them. This is the process that &get studied so extensively. As the human child develops, it builds up a lstock of mental patterns of control (“schemes”) that, in combination, control thought and action. The key word in the previous sentence is mmbinution, and in that word lies the greatest challenge for the psychology of the future. For in considering an extendible modelling system that uses composable patterns, several critical questions emerge. How can patterns of different levels of abstraction combine to model a specific event? What calls them forth and composes them appropriately? Remember that they are not to be strung together like beads on a chain but must be composed to form a structure that describes the various aspes!s of.the current situation, the molar and the m* leeular, the abstract and the concrete, at the same time. The understanding of a single sentence of language must surely involve the participation of bundreds, if not thousands, of computational schemes, some innate and some learned. How can they all be pulled in at the right time and cooperate smotithly? This is a p;oblem in system architecture, the solution of which may some day tell us how it is possible for a being to be truly intelligent. Closely linked to this issue of architecture is the problem of cognitive growth, or learning as it used to be called. How can a learning process create new patterns
38
A. S. Bregmm
for modelling previously uncaptured aspects of reality and add them to the current set without playing havoc with the already-existing system? I believe that the critical questions that I have mentioned, those concerning the combinability and [extendibility of the set of ideals, will have to serve as the focus for any seriolus future theory of learning. Indeed, one could speculate that these two problems are so closely linked that it is the same design feature in the human brain that both permits it to combine ideals as freely as it does and to add new ones to its repertoire. I believe that a theory of learning that addressed these issues could serve :is a point of unification of psychological theory. Perceptual theory, for example, would be seen to fit in smoothly with learning theory because both would be concerned withideals-perceptual theory with how we use them to model situations and learning theory with how we acquire and modify them. An extended version of Piaget’s theory would probably be closest to the mark. Paradoxically, it seems that despite the fact that Chomsky has identified himself with the nativist tradition in psychology, his most important influence may ultimately lie in having provoked the development of an adequate theory of learning. References Bregman, A. S. (1977) Perception and behavior as compositions of ideals. &g. Psychd. 9, 250-292. Bregman, A. S. (1978) The formation of auditory streams. In Requin, J. (ed.), Attention and Performance VII. Hihi~e, NJ, Launexxe Erlbaum Associates. Bregman, A.-S. (1981) Ask& the “what for” question in auditory perception. In Kubovy, M. and Pomerantz. 3. R. (eds 1,PerceptualOtguniwion Hillsdale, NJ, Lawrence ErIbaum Associates. Chomsky, N. (1965) Aspects ofthe Theory ofSyntax. Cambridge,Mass., MIT Press. Skinner, B. F. (1957) VerbalBehmior. New York, Appleton-Century-Crofts.
Cognition, 10 (1981) 39-52 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
39
An approach to Universal Grammar and the mental representationof language JOAN BRESNAN* Massachusetts Institute of Technology
I. The need for cognitive constraints on 80guistic representation Despite the exotic variety of the world’s languages, any normal child is capable of mastering any language. This fact suggests that there is some universal system for mentally representing natural language. The aim of Universal Grammar is to discover this system, One approach to Universal Grammar is to define formal symbolic systems that describe properties of various natural languages and then attempt to abstract a common formal structure from these systems. This is the formaldescriptive approach of generative linguistics, and it has fundamentally advanced our understanding cf the structure of na.tural language. Nevertheless, this approach by itself can yield only a limited understanding of how languages are mentully represented. The reason is that, apart from the requirement Cat any proposed fomd system (or generative grammar) bear a descriptive relation to language users’ knowledge of the language, there are virtually no cognitive constraints imposed on these systems. But the same knowledge can be represented in descriptively equivalent ways by symbolic systems whose design constraints would serve very different cognitive processes, ive a simple example, a symbolic system which is designed to minimize the number of primitive operations may be equivalent to one which employs more primitives but minimizes the length of derivations of well-formed formulae. Thus, the propositional calculus which utilizes the sole connective ‘4’ which means ‘neither nor’ is equivalent to the calculus that is based on the more familiar logical connectives ‘-‘, ‘v’, ‘+‘, etc., but
*This articleis based in partupon &wk supported by the National Science Foundation under Grant No. BNS 80-14730, and summarizes ideas that have developed in the close collaboration of members of our resemh group: Ron Kaplan, Marilyn Ford, Jane Grimshaw, Kris Halvorsen, Steve Pinker. Reprint requests shotid be sent to Joan Bresmw, Department of Linguisticsand Philosophy, MIT, Can&&e, Mass. 02139, U.S.A.
40
Joan Bresnan
in the former ‘p+q’ is expressed as ((pJp)Jq)J ((pJp)Jq)). Which design the constraint- ;-eduction of primitives or compression of formulae-is right one? The answer of course depends upon the purposes for which the symbolic system is used. ln the same way, descriptively equivalent linguistic systems may satisfy design constraints which would be optimally suited to very different cognitive processes (see Bresnan, in press).
2. The link between linguistic representations and cogni!ive processes To take a deeper approach to Universal Grammar as a theory of mental representation, we must recognize that the fundamental problems of Universal Grammar are problems of understanding the natural informationprocessing mechanisms of humans. The solution of a problem of natural ;nformation processing requires the integration of the levels of computational theory, algorithm, and process, as the work of Marr and his coworkers shows* (Marr and Nisihara, 1978). The first step is to identify the inforrnat-ion-processing problem to be solved. In general, this takes the form of specifying the desired (goal) representation and the given data (o I input representation) from which the goal is derived by some unknown process. The second step is to show theeretically how a reliable representation can be derived from the available data. Here the aim is to determine the natural constraints that guarantee the existence of a unique solution to the problem. The computational thepry developed at this step goes beyond the formaldescriptive approach by showing how constraints on the zeal representation are related ta the nature of the computation that theoretically derives the representation. The computational theory explains w/z) the representation has the form that it has in terms of the solution, to an information-processing problem. The third step is to design a particular algorithm that correctly interprets the available input information. It is this step which connects the level of theoretical representation with that of process, making it possible to study experimentally how properties of the representatiorrs of knowledge are related to properties of the mental processes that construct, maintain, and interpret those representations. The fourth step is to test whether the natural information-processing system uses the particular algorithm by comparing the behavior of the implemented algorithm with the observable processes of
*The recentdeath of David Marris a greatloss to all workers in cognitive science,but his ideas and his memorywill continue to inspireus.
An approwh to UniversalGmmmar
41
the natural system. If the comparison falls, a new algorithm must be designed and the testing continued. All of the major problems in the mental representation of language can be approached in this way, each requiring a computational theory of some natural human capability for information-processing. Thus, broadly speaking (for each problem decomposes into a set of more tractable subproblems), the problem of language acquisition is to specify a learning function that, given a universal store of initial (perhaps innate) knowledge, maps the primary data available to the child language learner onto a representation G of mature knowledge of the language. Similarly, the problem of language comprehension is to specify a decoding function that, using the language-user’s stored knowledge of a language, call it G again, maps auditory or visual data from the language onto mental representations of meanings. Likewise, the production problem is to find the inverse mapping, which, given linguistic knowledge G, encodes meanings into visual or auditory representations. Even the descriptive problem of generative linguistics is in principle no different: the problem is to specify a function that, given the grammar G of a language L, can associate all strings over the vocabulary of L with the judgments of wellformedness, ambiguity, systematic paraphrase, and the like, of those who know the language. 3. The competence hypothesis :JSa source of constraint It is evident t!lat the na?ural constraints on the mental representation of linguistic knowledge must derive from all of the interacting informationprocessing sy4terns that support language use. The so-called ‘competence hypothesis’ e: Eresses this fact by asserting that the same representation of grammatical ‘knowledge G is included in all of these systems (Bresnan, 1978; Kaplan, 1975; Chomsky, 1965). The competence hypothesis means no more than that the stored knowledge of a language acquired by the language learner is the same knowledge unconsciously utilized by the adult in comprehending and producing language, and that this linguistic knowledge which permits people to speak, recognize, and. understand their language is the same knowledge that enables them to tell the linguist whether any given utterance belongs to their language. The competence hypothesis is a powerful unifying assumption for theoretical linguistics and psycholinguistics, It implies that linguistic theory and psycholinguistic theory are interdependent, and that the problem of how ; knowledge of language is mentally represented cannot be understood in the absence of any theory of the mental processes required to construct, maintain, and interpret proposed representations of linguistic knowledge.
42
Joan Bresnan
After Miller’s (1962, 1964) pioneering work in support of the competence hypothesis, experimental psycholinguistics began to accumulate substantial evidence showing that the form of linguistic knowledge representation giverr by standard transformational generative grammars is not employed in angi straightforward fashion in sentence perception, production, and language acquisition (Fodar, Bever and Garrett, 1974; Levelt, 1974). Since more successful ways of employing transformational grammars in psycholinguistic theories have not been forthcoming, a number of researchers have concluded from these results that the competence hypothesis cannot be maintained. This conclusion, however, is unwarranted: from such results WC can equally well conclude that the form in which linguistic knowledge is mentally represented- G-cannot be identified with transfol.llatio& grammars (Bresnan, 1978). To conclude that the competence hypothesis itself is incorrect would be to relinquish the major source of cognitive constraint on representations of linguistic knowledge, thus losing any real possibility of explaining the interactions among the various linguistic information-processing systems, and multiplying theories of the mental representation of language without constraint. To the extent that these psycholinguistic considerations are well-founded, they indicate that a new form of linguistic knowledge representation-a new theory of Universal Grammar---is needed. The Sam<:conclusion can be drawn from theoretical consideration of a problem central to Universal Grammar-the syntactic mapping problem. 4. The syntactic mapping problem The syntactic mapping problem is one of the major information-processing
problems of Universal Grammar. It is the problem of finding for every natural language a Imapping which will associate each sentence of the language with a representation of its grammatical relations. (The term ‘grammatical relations’ is used neutrally to refer to the associations-between the surface word and phrase configurations and the semantic predicateargument structures of a sentence.) The syntactic mapping problem is exm tremely difficult -first, because of the complex, many-to-many relations between the sentences of any natural language and their grammatical relations, and second, because of the radical variations in surface form across languages. However, the competence hypothesis imposes important natural constraints on the solution to the problem, and these constraints limit the forms in which knowledge of language can be represented. Five constraints on the mapping can be identified, the first two of which are presupposed by all theories of generative grammar. The five are cratNtpr
of the mapping are theoreti&ly infinite), finite a finite capacity for the knowledge representations ), reli&Wy (the mapping is effectively computable), mmaticai relations that the mapping derives a sentence must be directly included in the the mapping derives from the entire sentence, o:h prior or subsequent segments), and untveras a universal procedure for constructing repelations). Each of these constraints has a ~1 motivation: for example, order-free composition is motivated by our fluent and relatively effortless ability to interpret sentence fragments; universali@ is motivated by the interaction of language acquisition and language compwhension. These constraints on solutions to the syntactic mapping problem impose important limitations on the possible forms of syntactic knowledge representation, ruling out many possible systems of grammar -even apparently descriptively adequate ones such as transformational grammar -as systems of the mental representation of language (see Bresnan, in press). A soIutiort to the syntactic mapping problem does exist. For any language L, let MPt&e G to be a lexical functional grammar for the language, and R (the represerrtation of grammatical relations) to be the set of functional structures of ahe language as defined by G. Then a map m: (L, Gj -) R exists which satisfier all of the given constra!nts. This result ir, based on the mathematical characterization of lexical fun ,tional grammars by Kaplan and Bresnaxl (1980). It provides the foundations of a computational theory (in the sense given above) for investigating the mental processes that construct representations of syntactic structure.
(the domain and capacity (there is
5. Lmhl
functional
mmar
A lexical functional grammar. provides each sentence of a language with two structures-one representing its surface form (the constituent structure), the other representing its grammatical relations (the functional structure). The functional structure represents grammatical relations in a universal format, abstracting away from language-particular and typological charaeteristics of the surface forms of sentences. The universal format is achieved by using grammatical functions (SUBJECT, OBJECT, etc.) rather than languageparticular phrase structure configurations or morphological case to specify the associations between semantic predicate argument structures and the surface forms of sentences. The universality of grammatical functions is a fundamentally important contribution of research in relational grammar
(Perlmuttel, and Postal, 1977.),which has been adopted even in some current versions of transformational grammar (Chomsky, 1980). However, the role of grammatical fi.1qdions in lexical functional grammar differs from both of these in essential ways. In the theory of lexical functional grammar, the predicate argument structures of lexical items are represented independently of their syntactic contexts-I features as functions of a fixed number of grammatically interpretable arguments. A mapping between predicate argument structure and syntactic constituent structure is specified by means of grammatical functions. These are assigned to surface phrase structure positions by syntactic encoding functions and to predicate argument structure positions by lexical encodi.rg functilons. A predicate argument structure with grammatical functions specified is called a lexicalform. How these terms apply to a simple English sentence is illustrated in Figure 1. Figure 1.
tb
------------------
the baby
_-_---em.----
An approach to Univem! Cmmmar
45
In Figure 1, the information above the dotted line is provided by the syntactic component of a generative grammar for English, and that below the dotted line is provided by the lexical component. Thus the lexical form for the verb hand has a triadic predicate argument structure whose three arguments may be identified with the thematic roles SOURCE, THEME, and GOAL (Gruber, 1965; Jackendoff, 1976). Arguments 1, 2, and 3 of this predicate argument structure have been assigned the respective grammatical functions SUBJ(ect), OBJ(ect), and OBL(ique)ooAL by lexical rules. The syntactic component generates the surface phrase structure tree for -r;‘,,d handed a toy to the baby and identifies the NBSdominating Fred, a ?oy, and the baby as SUBJ, OBJ, and OBL~~AL, respectively. In Figure 2, the verb hand has a different lexical form, produced by assigning a different set of grammatical functions to the same predicate arguFigure 2.
phrase st_ructure surface grammatical functions
handed
lexical form for hand
hand ((SUBJ), (OBJ2), (OBJ)) _
------I
(SUBJ)
(OBJ2)
--_
-
-----
(OBJ)
Iland ( (SOURCE) (THEME) (GOAL)
------
-----
lexicalassignmentof grammatical fur-:tion predicate argument structure
46
Joan .Bwsnan
ment structure as before. As shown in Ffgdre 2, OBJ is now assigrred to argument 3 and 0BJ2 (second object) is assigned to argument 2. Note that the position of grammatical function symbols in a lexical form is completely independent of the left-to-right order of constituents in phrase structure. When the structural grammatical functions are matcheal with the lexical grammatical functions in Figures 1 and 2, it is evident that Fred, a toy, and the baby correspond to the same ‘logical’ arguments in the predicate argument structure of hand. Both syntactic structures can be base-generated, and the seeming transformational relationship between thelm can be expressed instead as a relationship between the lexical forms for ikund. (Details of the dative alternation- and its interaction with passivization and other rules are discussed in work by Bresnan [ 19801.) Thus the verb hand has several lexical forms, and general conditions on functional structures ensure that the- correct lexical form is paired with the appropriate syntactic structures ,Xaplan and Bresnan, 1980). The functional structure of a sentence assembles the relations between the syntactically encoded functions and the lexical forms; thus it is the functional stru ture that represents the associations between surface form and semantic p,.zdicate argur:ient structure, and these associations constitute the grammatical relations of the sentence. Figure 3 shows the functional structures for the examples of Figures 1 and 2. Note the very direct relation between the surface forms of these examples and their respective functional structures. Figure 3.
Functional structure of
Functionalstructure of
Fred handed a toy to the baby
Red handed the baby a toy
1
SUBJ
SUBJ
rENSE Past
TENSE Past
PRED ‘hand((SUBJ), (OBJ), (CJB&,&)
PRED ‘hand((SWBJ), (OBJ2), (OBJ))’
GBJ
1
SPEC ‘a’ PRED ‘toy’ c
oB&ML _
[
OBJ
1
SPEC ‘the’ PRED ‘baby’
PRED ‘Fred’
~::tiyj
0852 kfWy,
]
An approachto Universal Grammar
47
The directness of this relation reflects a fundamental property of the lexicalfunctional theory of grammatical representation: only lexical rules can alter grammatical relations; al! syntactic rules must preserve functionassignments (Jackendoff, 1976; Kaplan and Bresnan, 1980). It is this property-the principle of direct syntactic encoding-which ensures that lexical functional grammars satisfy the order-free composition constraint; for now surface syntactic forms can be specified entirely by context-free gremmars (or recursive transtiion networks (Woods, 1970) and the functional structures can be built up by a simple ‘additive’ operation that applies directPy to surface forms (Kaplan and Bresnan, 1980). The functional construction process is effectively computable, satisfying the reliability constraint. Similarly, it is the universal format of functional structures that ensures the universality constraint. The dative alternation shown in Figures 1 and 2 clearly illustrates these principles. Since in Figure 1 the OlBJCject)NP is associated with the THEME argument, while in Figure 2 the OBJ(ect) NP is associated with the GOAL argument, the grammatical relations of the sentences differ. Hence the dative alternation must be effected by a lexical, rather than a syntactic, rule. In contrast, the examples Fred gave a toy to the baby and Fred gave to the baby a toy YIOno; differ in their grammatical relations: in both examples, the OBJ(ect) NP is associated with the THEME argument and the prepositional phrase is associated with the GOAL argument. Hence the different orders of the phrases must be effected by a syntactic rule (In this case, one which simply specifies alternative sequences of constituents of the verb phrase). 6. Alternative theories of Universal Grammar It is evident from the principle of direct syntactic encoding that any rule of grammar which changes the grammatical functions of constituents must be a lexical rule. Moreover, any such rule will have a universal characterization which reveals its invariant form across languages. l%is follows because grammatical functions are independent of language-particular realizations in terms of syntactic structure or morphological case. The lexical functional theory thus differs in essential ways from transformational’ theories (Chomsky, 1965, 1980) from other structuralist theories (Gazdar, in press; Peters, in press), and from Relational Grammar (Perlmutter and Postal, 19775. AI1 versions of transformational grammar share the fundamental representational principle that at some (‘deep’) level of representation, there is a one-
48
Joan Bresnan
toone correspondence between the predicate argument structure (or thematic role structure) of a sentence and its phrasal structure. For example, Chomsky (1980) refers to this as the ‘theta criterion’. The associations betweer predicate argument structure and surface form must then be effected by operations on phrase structure representations (such as syntactic transformations or equivalent structure-dependent rclles). But it is a fact that surface phrasal structures of natural language sentences do not bear an isomorphic relation to their predicate argument structures: for example, give has the same predicate argument structure in Figures 1 and 2, surprised has the same predicate argulment structure in John was surprised at Mary’s idea and Mary’s idea surprised John, and prevent has the same predicate argument structure in You can ‘t prevent it from raining and You can’t prevent it; raining; yet there is no 3omorphy between the surface phrasal structures in each pair of examples. Consequently, the syntactic mapping between predrcate argument structure and surface forms must faii to preserve grammatical function assignments, violating the principle of direct syntactic encoding. This result can only be avoided by giving up the basic representational prmciple of transformational grammar and permitting multiple lexical correspondences between predicate argument structure (or thematic role structure) and phrasal structure, as in the lexical functional theory of grammar. The idea that grammatical functions are defined and derived on phrasestructural configurations such as deep structure is a fundamentally structuralist conception of grammatical relations which has been preserved in transformational grammar (ChJmsky, 1965, 1980). While the structural% conception seems descriptively adequate for languages like English, it clearly fails to provide a unified theory of grammatical relations for languages having radically different surface forms. For example, it has been shown that the structuralist representations of transformational theory are insufficiently abstract to characterize the universal attributes of rules like passivization (Bresnan, 1980; Mohanan, 1980; Perlmutter and Postal, 1977). RecenL work within the lexical functional theorv of grammar shows that at a suitably abstract level of grammatical representation, two languages that differ radically in their phrase structures exhibit identical lexical encoding, and that the strikingly different syntactic manifestations of these rules follow from a simple difference in the syntactic enccding function for each language. Moreover, this work shows that the appearance of ‘NP movement’ in constru:tiorr; like the. passive is an illusion created by the particular properties of the syntactic encoding functions for ‘configurational’ languages like English. Far from being the reflection of a universal transformational component of grammar, as many linguists currently maintain, the pheno-
An upproachto UniversalGrammar
49
menal characteristics of ‘NP movement’ are detivative of more abstract priniples of mmatical representation. Rulational Grammar, like lexical functional grammar, rejects the structuralist conception of grammatical relations and provides a universal theory of function-dependent processes such as the dative alternation, passivization, and causativization. But unlike the lexical-functional theory, Relational Grammar has assumed that there is a one-to-one correspondence between predicate aguments and an initial assigment of grammatical functions. These predicate arguments are mapped onto surface constituents by a sequence of ‘strata’ which syntactically modify the initial assignment of grammatical functions until a final assignment of grammatical functions is produced. The final grammatical function assignment is then mapped onto surface strings by linearization rules. Because of the functionchanging character of the syntactic mapping, Relational Grammar violates the principle of direct syntactic encoding, and so differs from the Iexical-functional theory in the decomposition of grammars into lexical and syntactic rules. But there is evidence from the interactions of passivized verbs with rules of word formation that passivization is indeed a lexical, not a syntactic, process (Bresnan, 19%)). Generalized phrase structures hPl.ri:been developed by Gazdar, Peters, and others. In their use of a single level of constituent-structure representation, these theories of grammar resemble the lexical-functional theory and differ from both transformational theories and Relational Grammar. But lexical functional theory contrasts with generalized phrase structure theories in its claims that Yunctional structure, not constituent structure, is semantically interpreted (Haltorsen, in preparation) and that there are universal and language-particular generalizations that are captured only in function-iiependent terms (Bresnan, in press). The fact that there are natural languages in which sentences lack virtually all constituency relations poses a severe problem for generalized phrase structure grammars as universal theories of grammar. For example, the Australian idnguage Ngarluma (Simpson, 1980) exhibits to an extreme degree the lack of phrase structure that Hale has identified as a typological property ori natwAx language (Hale, 1979). In Ngarluma, the sentence consists of a stt;ing of words without phrasal structure; nevertheless, the perceived grammatical relations of the sentence in Ngarluma have the: same kind of functional structure as those in English. In particular, the Ngarluma sentence has grammatical subjects and objects and a rich system of predicative : fa adjunct modifiers, all of which may be composed of information derived from several words which are discontinuous in the string; it differs from a language like English in that these functional units do not correspond to phrase structure constituents-they are syntactically encoded in terms of case inflections rather than phrasal structure.
50
Joan &man
7. Toward a’unified theory of the psychology of language
The design constraints embodied in theories of Universal Grammarhave profound consequences for theories of the mental representation of language. Research in the theory of lexical functional grammar has begun explicitly ‘to investigate these consequences. Pinker (1980), arguing that lexical accounts of acquisition have certain advantages over previous syntactic theories, has proved that lexical functional grammars are learnable in principle, and has developed principles underlying a new theory of language acquisition based on lexical functional grammars. Ford, Bresnan, and Kaplan (198 1) have constructed and motivated a competence-based theory of syntactic closure which directly incorporates lexical functional grammars as representations of linguistic knowledge. A particularly interesting and important research problem in the development of a theory of sentence perception has been to explain the structural biases shown in structurally ambiguous sentences, since the effects observed have been assumed to reflect the operating principles of the human parsing mechanism very directly. The study by Ford, Bresnan, and Kaplan examines syntactic bias effects and shows that they are a joint function of (i) the linguistic rules which define the structures of sentences, (ii) the predicate argument structures and grammatical functions of lexical items, and (iii) a well-defined interaction between rule-driven and data-driven analysis procedures. Although relatively little psycholinguistic research on adult language behavior has been concerned with how speech is produced, Ford (1980) has completed an experimenti study on this problem. Together with the results obtained by Ford and Holmes (1978), the findings provide very strong evidencle that sentence production proceeds by the successive planning and uttering of segments of a sentence called ‘basic clauses”. Raising the question of what representation of the meaningful relations underlying sentences would permit sentences to be planned basic clause by basic clause, Ford shows that the required type of representation is one in which for each basic clause, there is a representational unit which encodes the meaningful matical relations for the surface items in that clause, such as the surface subject, object and verb. Ford gives evidence that these representations cannot be analogous to the representations of underlying relations in various versions of trzmsformational grammar, and shows that the lexical functional theory of grammar does represent underlying relations in a way that would permit sentence planning to proceed basic clause by basic clause. Because of the computational theory provided by lexical functional grammars, it has become feasible to consider the design of explicit algorithms for
acqu entations of syntactic lcnowllevel of th representation with that of 1s will make it possible to study experimentally how properties of the ~p~~~tati~n of syntactic knowledge are related to properties sf the mental processes that construct and interpret those representatisns. and
References Bresnan,J. (1978) A Realistic Transformational Grammar. In M. Halle, J. Brescan, and G. Miller, (eds.1, ~f?~g~ubtic Theory 4nd YsYchc~ogicuII&&y. Cambridge, Mass., The MIT Press, Bresnan, J. (1980) The Passive in Lexical Theory. Occ&ortul f4p4r #7, The Center for Cognitive Science, MT; also to appear in J. Bresnan (ed.), The Menfd Represenfution of Gra~muficul R&ions. Cambridge.,Mass.,The MlT Press. Bresnan, 1. (In press)UniversalGrammar and the Mental Representation of Language. Introduction in J. Bresnan, (ed.), T%re Menfd Represenfufion of Gn;mmuticui Relations, Cambridge, Mass., The MIT Prea. ChOmSkY, N. (1965; Aspects offhc Theory ofSynfcuc. Cambridge, Mass.,The MITPress. Chomsky, N. (1980) On the Representation of Form and Function. Presented at the C.N.R.S.Conference at Royasmont, France, June 1980. Fodor, 4. A., Bever, T. and Garrett, M. (1974) ThePsychoioay oflunguage. N.w York, McGraw-Hill. Ford, J. M., Bmsnan,1. and Kaplan, R. (1981) A Competence-BasedTheory of Syntactic Closure. Ckc4sfund Paper #II The Center for Cognitive Science, MIT; also to appear in J. Bresnan biut.A Y%eM?fffd Ropmenfation of Gmmaficai Rd4tkm. Cambridge, Mass., The MIT Press. Ford, M. (1980) Sentence PkmningUnita: Implicationsfor the Speaker’s Representation of Meaning firI Relations UadarSving Santencee. Occasionirl Pup&v #2, The Center &Jr CognitiveScience, HIT; also to appearin J. Bresnan(ed.), The MeMa Repraenfufion of Grammaticd Relations. a,, nte MITPrese. Ford, M., and Holmes,v, (197$) Plnnnhg.Units in Syntax and Sentence Production. Cog., 6, W-53. Gazdar, 0. (In press) E&rase Structure Grammar.In P. Jacobson anc, G. Pullurn (cds.), The Nature of 8y~factk Re~~~~tff~~, London, Groom Helm, Gruber, 1. (1965) ~~~~~s In [email protected]/ R&,imr, MITdactoral dissertation, reprinted by the lndiana
UnivoreityLinghties Club, Bloomington,Indiana. Hale, K. (1979) On the PosWwa crfW&d in o Typdogy sf the &se. Department of Lin$ustics and philosophy, MIT. Mamser~pt avaIlablefrom the Indian2 WversjY Linguistics CiuL, 1981, Bloomington, Indiana. Hakorson, P.-K. (In preparation) An lntcrprettive Procedurefor Functional Structures. To appear as an Q&WOWJ P@erc The Centerfor Cognitive Sdence, MlT. Jackendoff, W. (1976) Toward an Explanatory Semantic Representation.Linguistic Inquiry, 7,
89-150. ~aphn, R. (1979) On ptacca yodels for Sentence Analysis. In D. Norman and D. Kumebrt (eds.), i?xplrvn%rc,vIn Cognition. !&IIFrancisco, W. H. Freeman and Co. Kaplan, R. and B;esnan, J, (1980) LexichFunctional Grammar: A Formal System fcr Grammatical Representation. &ca~iod Puper #f3, The Center for CognitiveScA-wc, ET, Jlso lo *PpeN
52
Joan &esnan
in J. Bresnan (ed.), The Mental Representation of Grammaticalkelations. Cambridge, Mass., The MIT Press. Level& W. J. M. (1974) Fotwtal Grammarsin Linguisticsand Psycholinguistics,3 ~01s. The Hague, Mo;lton. Marr, D. and Nislhara, K. (1978) Visual Information Processing: ArtiWcial Intelligence and the Sensorium of Sight. Technol. Rev., 82* (1). Miller, G. A. (1962) Some Psychological Studies of Grammar.&. Psychol., 2 7,748-762. Miller, G. A. and McKean, K. (1964) A Chronometric Study of Some Relations between Sentences. Q. J. Exper. Psychol,, 16, 297-308. Mohanar:, K. i . (1980) Grammatical Relations and Clause Structure in Malayalam. Manuscript, Departm.ent of Linguistics a,.3 Philosophy, MIT; to appear in J. Bresnan (ed.), The Men&J RqreseMation of GrammaticalRelations.Cambridge, Mass., The MIT Press. Perlmutter, D. and Postal, P. (1977) Toward a Universal Characterization of Passivization. In Proceedings of the Third Annual Meeting of the Berkeley Linguistics Society, Berkeley, California. Peters, S. (In [press) Definitions of Linked-Tree Grammars. Technical Report, The Cognitive Science Center of the University of Texas at Austin, Texas. Pinker, S. (1980) A Theory of the Acquisition of Lexical-Interpretive Grammars. OccasionalPaper #6, The Center for Cognitive Science, MIT; also to appear in J. Rresnan (ed.), The Mental Representationof GrammaticalRelations,Cambridge, Mass., The MIT Press. Simpson, J. (1980) Ngarluma as a W* Language. Manuscript, Department of Linguistics and Philosophy, MIT. Woods, W. (1970) Transition Network Grammars for Natural Language Analysis. Communicationsof the ACM, 1X 591-606.
Cbgnition, 10 (1981) 53-58 @ Elsevier Sequoia &A., Lausanne - Printed in The Netherlands
53
ive and control processes DONALI) E. BROADEWdT+ Uniwmity of Oxford
If we rpp;.oach a depressed patient and ask him what he Z.sthinking about, he is likely to report that he is remembering some episooe of failure in the past, or contemplating the hostile expression of a passer-by. Yet the patient may be sitting at his ease on a sunny day, with many more pleasant sights and memories which he is ignoring; and he himself may bitterly regret his inability to select these other and more congenial topics. This e
of Oxford, South Parks Road,
of which more than half the N features were present rather than absent. It would then be (quite a major operation to decide about each candidate whether or not it beloqged to the desired sub-set. In the worst cases every single feature would seed to be checked before one could decide. There are many intermediate methods of partitioning; for instance, one could say that two features rather than one were relevant to the selection, but that a.dmission to the subset required that the two features were either both present w both absent. Man;4 of the properties of different methods of partitioning an ensemble have of course been analyzed in a classic book by Garner ( 1962). There is no ‘best’ way; it depends on the purpose for which the partitioning was made. If for instazrce one is selecting a sub-set from amongst physical objects, and there b some cost attached to observing each feature of each object, then there are obvious advantages for the first c f the methods suggested above. If we loak at one feature only for each object, fewer observations will need to be made. But suppose there is some error in the observations. Suppose you know that, for each object examined, one of the features detected will in fact have been absent, and one of those reported as absent will in fact have been present. In that case, selecting a sub-set on the basis of one feature would be perilous; some objects would be wrongly included, and some which should have been selected will be missed. It would be much more reliable to use one of the other methads of partitioning; foi instance, in the case mentioned the partitioning based on the detected presence of ILlore than half the features would give no misclassification at all. In general, selection on the basis of a single feature will economise on processing, while selection on the basis of a combination of features will mininiise the risk of misclassification. Now, when we perceive we normally select some part of the surroundings. Logically, we could do this by any of the methods of partitioning mentioned already. As an example of one extreme, let us take a person who is listening to any wards which are spoken from the left-hand side of the room, whatever those words may be; at the other extreme, consider a person who is listening to any words which are animal names, whatever direction they come from. These are single-feature and’multi-feature methods of selection; they are also of course the mechanisms of ‘filtering’ and ‘pigeon-holing’ distinguished by Broadbent I1 97 1). The bulk of studies on selective attention continue to ignore the difference between different methods of specifying target sub-sets (at the cost of terrible confusion in the literature). Where the distinction has been preserved in more recent studies,, however, the results have gone on showing that one kind of selection is very different from the other (Corcoran and Jackson, 1977; Keren, 1976). A particularly clear case is a study by Treisman and Gelade (1980) in which people trying to select a letter R fmm amongst other tiual letters were able to do so (after practice) at high speed if the ir-
Selective
and control processes
55
relevant letters were Ps and Bs, which have no diagonal line as the R does: if however the irrelevant letters were Ps and Qs, so constructed that no sing’,e feature of the R distinguished it from the others, then selection was much more difficult, even after practice. The method of selection employed by the person has a large effect on the degree of success achieved. Thus far,we have thought only of selection in perception.But memory raises the same problem ; people’ carry’within them a large’ number of representations of past events. At any one time we recall only a sub-set of these events, leaving the others to show themselves on some other occasion. Logically we could use the same range of methods as in the case of perception. Selectic:a. might be of events possessing some single feature. Somebody trying to remember a list of words might first select words which were names of precious stones, then those which were parts of the body, then foods; rather as a librarian might collect books from different parts of the library. Indeed, recall does tend to cluster in this way (Bousfield, 1953), and it is wellestablished th.at people can recall much better if they are given material structured in a fashion resembling the hierarchical plan of a library. (Bower et al., 1969). On the other hand, recall might be on a multi-feature basis. In fact, material structured in tables or matrices is on avera,ge just as easy to learn as hierarchic material is (Broadbent, Cooper and Broadbent, 1978). When individuals are given the chance to structure memory material as they choose (Broadbent and Broadbent, 1978), they show a range of performance from near-hierarchical to nearorthogonal allocation of dimensions. It seems likely that selection in memory, as in perception, can make use of several of the logically posrtible methods and is not confined to one. This certainly matters for specialists in memory, but why should it matter more generally in cognition? During the past decade computable models have been produced for those human actions which run off in changing, unstereotyped, sequences; such as the solving of puzzles (Newell and Simon, 1972) the seriation and class inclusion tasks of Piaget (Young, 1978) or the processes of medical diagnosis (Fox, 1980). This has been done, not by unreal rigid sequences of sub-routines, but by supposing that the action taken at any instant depends on the conditions present at that time. The system consists then of a collection of rules relating conditions to actions, a ‘Production System”. The conditions being examined include the external situation, but also the current state of short-term or working memory. What happens when the system runs depends therefore on locating one of the rules (productions) in longer-term store which matches the conditions holding at the senses and in temporw memory. Therefore, the probability of a certain action being taken will depend on the method of search being used for conditions on the one hand and for rules on the other. This is true even if the conditions of a rule are in fact sat-
56
D. E. Bmdbent
isfied; if the rule, or some of its conditions, are not selected before decision favours some other rule, it might as well not be in the repertoire. The way people act in dynamic sequential situations raises again all the problems of selection. Failures of that selection could explain why people make mistakes which, in another sense, they know perfectly well they should not make; and why they seem incapable of handling problems that are obvious to an outside observer. So there are theoretical reasons for expectir:g parallels between perception and memory; and for observing what people do in situations which need a sequence of actions that runs off in a different order depending on the circumstances. Controlling a model of an economic system is one example; others include managing a computer model of a business, or interacting with a sirnulated human personality. Three of the disparate fea.tures of my programms begin therefore to make sense. What however of the fourth? Why should these interests lead one to interview welders, titters, or electricity sub-station at-
tendants? As we have seen, people are not tied to one form of selection either in perception or in memory. They can do different things in different situations. Nevertheless, in unstructured situations they prefer one or another technique. One possible line of inquiry is the generahty of this tendency, within an individual. If however there are preferred individual biases towards one method of selection rather than another, then one would like to relate those biases to some other feature of the individual. How about using the relative frequency of various information-processing tasks in the person’s experience? It might be that the need to employ one method of selection frequently might cause it to be used preferentially when other factors were equal; or it might be that people who had a pre-existing bias tended to seek out tasks which required such a bias. Either way we might find that people whose experience differ in a major way will show different strategies. Where is the greatest difference in the information-processing requirements of individuals? A good candidate is in bis work. It is easy to find1one man who spends eight hours a day placi,ng 5. piece of metal in position, operating a press, and then repeating the ope:ration every ten seconds. It’is also easy to find another man who spends three weeks moving a unique product through a series of different operations until it is finished, and then never repeats the same sequence. Do the two men differ in the frequency or ‘kind of their minor disorders of attention, in slips of memory or of action? Goals and ReaIism
By now the ultimate goal of the programme is clear. Ideally, one would Iike to reach some definitive statement of the form ‘people who perform unchang-
Selective
and control process 2s
57
ging sequences of operations at work tend to use filtering in perceptual tasks, to do well :at hierarchically structured memory tasks, to do badly at tasks which demand a different approach on every trial; and in every day life, they forget things when distracted’. Such a hope is of course simplistic. It is even mlikely that I shall achieve a more modest personal goal; to be able by 1991 (when I personally reach retirement) to take a microcomputer to a factory, and by appllying a perceptual measure be able to say that a man is under undesirable stress. What morC modest success might be realistic? The problem is that strategies of selection and control night a priori be related to each other, and to everyday life, in a number of different ways. FOJ instance, successful avoidance of cognitive failure in everyday life might need1 the same strategy in perception a*ld in memory; or it might need different ones (because the two situations place different demands on processing and on avoidance of error). It might be that the job affects the worker; or that certain workers pick certain jobs. Even more likely, people who are liable to cognitive failure may be ‘vulnerable’, likely to develop psychiatric symptoms if they enter stressful jobs but not otherwise. That in fact is the relationship best supported by the evidence we are getting (Broadbent, Cooper, FitzGerald and Parkes, 198 1). There are many points at which opposite assumptions might be true; and of cotrise the number of possible complete theories rises as a power function of the number or such points. That is why neither I, nor anybody else, is going to reach a con.,plete and valid theory of cognition within a decade. What we can certainly do is to narrow down the possibilities sharply, by launching empirical studies whic.h check out each assumption. The I%rststep has br LF POconfirm that the general area does indeed contain a problem. Mercifully, it turfis out that there is some connection between work, cognition, and stre:ss. We now kplow that different jobs show differences in the frequency of psychiatric symptoms; the results given by Broadbent (1981) are only representative of a number of other findings. Furthermore, it is also clear that th: number of pa;tchiatric symptoms is correlated with the frequency of cognitive failures in everyday life. (Broadben t, Cooper, FitzGerald and Parkes, 198 I ). We are now sure that there is a treasure hidden in this forest; it becomes worth systematic se:arch to narrow down which part of the forest contains it. That is, we must check some of the broader possibilities such as the generality of individual styles across tasks. Ten years should eliminate four or five questi,ons, and thus cut the number of possible theories by a factor of twenty or so. The great mistake would be to follow those who try to conjecture a cornplete theory, who fall again into the seductive belief of philosophy, that ‘you can’t play Twenty Questions with Nature and win’. This leads them to try to guess, in advance, the answers to all the detailed assumptions needed by the
58
D. E. hadbent
ultimate complete theory. Unfortunately, this is not a practicably strat for reasons I have often explained (e,g., Broadbent, 1958, pp. 307~313). It does not avoid playing Twenty Questions; it merely does so ine~~~i~ntly~and gives a very high probability of losing. The efficient way is to ask questions which maximise the information gained from the answers; and that way in turn requires that each experiment tells us as much if it fails as if it succeeds. The surest rule is to make the studies decide between practical alternatives for action (Broadbent, 1980). Of course, there is always the alternative of a true refusal to pIa)- with Nature at all; you can abandon the attempt to understand the relation between the jobs which men and women do, and their cognition. But that is a basic choice which I made long ago. References Boustield, W. A. (1953) The occurence of clustering in recall of randomly arranged associates.J. gen P~chol,, 49, 229-240. Bower, C. H., Clark, M. C., Lesgold, A. M. and Wiuzenz, G. (1969) Hierarchical retrieval schemes in recall of catcogorical word Bsts.J. verb. Learn. verb. Sehav., 8, X3-343. Broadbent, D. E. (1958)P&ce~tion and Communfcation.London, Bergamon Press. Broadbent, D. E. (1971) Decision and Stress. London, Academi- ?ress. Broadbent, D. E. (1980) The minimisation of models. In A. 3. Chapman and D. M. Jones (ads.), Mu&Is of man. Leicester, Britich PsychologicalSociety. Broadbent, D. E. (1981) Chronic effects from the physical nature of work. In B.CardeBand G. Johansson (eds.), Manand worMngIWe,London, Wiley. Broadbent, D. E., and Aston, B. (1978) Human control of a simulated economic system. &gonom., 21, 1035-1043. Broadbent, D. E. and Broadbent, M. H. P. (1978) The allocation of descriptor terms by individuals in a simulated retrieval system. hgonom., 21, 343-354. Broadbent, D. E. and Broadbent, M. H, P. (1980) Priming and the passive/activemodel of word recqnition. In R.S. Nickerson(ed.),Rrrentbfl ondPkvformance VIII.Hi&dale, N&LawrenceErlbeum. Broadbent, D. E. and Broadbent, M. H. P. (1981) Recgncy efl%cts in vieud memory. Qua J, exg. PsychoI., 33A, l-14. Broadbent, D. E., Cooper. P. J. and Broadbent, M. H. P. (1978) A cornpa; n ofhiwtchkalsnd ma6slx retrieval schemes in recall. J. l&p, Psychol: Hum. Learn. Mem., 4, 486-497. Broadbent, D, E., Cooper, P. J. FitzGerald,P.and Parkas, K. R. (In Press) The Co8nitive Failures Queae tionaire (CFQ) and its correlates.Mt. J. CYin, p$ydrsl. Corcoran, D.W.1. and Jackson, A. (1977) Basic processesand strate@esin visual search. In S. Dornic @cl.),Attention anrlPerf&nunce VI. Hi&dale, NJ, Lawrence Eclbaum. Fox, J. (1980) Makingde~tions under the influence of memory. Bychd Rev., 87,190-211. Garner,W.R.(1962) tlhcertdnty cmttd Sttwure as Psychdogfcd Concepts. New York, Wiley. Keren, G. (1976) Some considerations of two allegedkinds of selectiveattenti0n.J 6xg Rtychol Gea, 105,349-374. Newell, A. and Simon, H. A. (1972) Human problem soluing. Englewood CBffs, NJ, Prentice Hall, Treisman, A. M. and Gelade, G. A. (1980) A feature-integration theory of attention. 6bg. &y&i, 12, 97-136. Young, R. M. (1978) Strategiesand the structure of a cognitiveskilL In G. Underwood (ed.), &m&g&~ Of fnfcwmation&cessing. London, Academic Press.
Cognition, f 0 ( 1981) 59-64 @ ElsevierS.:quoia S.A., Lausanne - Printed in The Netherlands
59
Prospectsfor neuralinguistic theory
Studies Of aphasia and theories of la e-brain relationships have advanced ~~~sider~bly in the past ten years, t point where it seems possible to atpt 8 gfwer& statement of the goals of neurolinguistic theory and to suga ms;Tzch program directed towards these goals. In my view at least, the ificant steps towards the clarification of the nakre of neurolinguistic d towards the actual development of theories in this field, have resulted from the introduction of concepts from linguistics and psycholinguistics into the study of language disturbances. This is not to deny the significance of other approaches, but rath*r to suggest that work in other fields may be integrated into a theoretical framework suggested by current linguistic and psycholinguistic results. The goal of neurolinguistic theory is to relate terms in a theory of language structure, acqcidtion, and use, to ternIs in a theory of neural structure and function. This a,oal makes a number of philosophical assumptions which seem perfectly natural qd will not concern me here. It obviously requires that we specify theories of language structure and use, and of neural Etructure and function, in terms which are conceptually compatible (in the sense to be explaincd below) and which are related empirically. It leaves open the important possibility that our theories of both language and neural structure can be modified by what we discover about their interaction. Contemporary work m linguistics has moved towards a definition of the nature af the theory of language in ways which are directly relevant to the establishment of the dotnain of phenomena which neurolinguistic theory must addrew. The revolution’ in linguistics has defined a set of elements and opernixed in highly constrained ways, which constitures a thr,ory of grammar. A grammar can be seen, very roughly, as the representation of the language structures a mature language user has a; his disposal. Contemporary theories of grammar have provided tentative but highly specific hypotheses regarding universals of language, the nature of the stsbstantive elements and formal operations inherent in various components in the grammar, constraints upon these operations, parameters whose speeitication during development would lead deductively to a richly specified set of struc;:ures in a particular ‘R@print mqueds should be sent to: David Caplan, Division of Neurology, Ottawa Civic Hospital, Ot?awa, Ontatio, Qnada.
language, and wsll defined grammatical levels which play particular roles in mediating sound-meaning relationships. One important discovery has been that, essentially without exception, the structures and operations which appear in modern theories of grammar do not find counterparts in theories of other cognitive capacities, either in humans or in other species (Chomsky, 1980, and the references cited there), a finding which has led to the view that a grammar is a species- and domain-specific cognitive structure which interacts with other such structures to yield normal intellectual function, including actual language use (Klein, 1977). These linguistic discoveries have had the effect of focussing the attention of psychologists upon the question of the processing of linguistic structures, rather than upon the study of the conditions of their use. It has come to be appreciated that any theory of language use, if it is to describe and account for the riehness of ‘verbal behavior’, will not be constructed without attention to the nature of the language code itself. The psychology of language has, accordingly, shifted from the largely unsuccessful efforts of the 40’s and SO’s to approach language with such tools as S-R and Freudian psychology (to name two completely contrasting approaches which are similar in this one respect) to the task of identifying a much narrower aspect of the psychology of language, the characterization of the unconscious processes whereby structures of a grammar are attained in the acts of speaking, comprehending speech, reading, etc. A highly productive effort on the part of experimental psychologists, theoretical psychoiinguists, and workers in Artificial Intelligence, has sharpened a number of questions in this area, and we now have models of prolzesses such as speech production (Garrett, 1978), syntactic parsing (Fodor and Frazier, 1980) and word recognition (Morton, 1978)-to name but three areas-which, though as tentative as modern linguistic analyses, are quite detailed, and specific enough to give rise to empi.rically approachable contra versies. Both the chara&er of inquiry in linguistics and psycholinguistics and the specific hypotheses of T-G grammar regarding levels of representation in the ‘language code’ are relevant to the study of neural mechanisms underlying language and of organically occasioned disordems of language. Obviously, the discovery (I think the term is appropriate) of heretofore unrecognized levels of linguistic representation, distinctions among language elements, etc., provides a vocabulary for Ihe descri)ltion of a number of aphasic disorder:s. We now have, for instance, a charact&zation of agrammatism in language-universal terms (Kean, 19803, alternate analyses of the generation of phonemic paraphasias (Lecours and Lhermitte, 1969; Schnitzer, 19’?2), hypotheses as to the nature of the linguistic representations which can be attained by the right hemisphere (Zai;del, 1978; Dennis and Whitaker, 1976),
Prospects for neurolinguistictheo,y
61
and other such studies; the same is true with respect to the development of models of language pathology based upon modern psycholinguistic concepts (Bradley et al., 1980; Coltheart et al., 1980). We are entitled to hope that yet other aphasic phenomena, such as paragrammatism, may come to be analyzed in such terms, and to expect that the study of diflkrent neurol.2gically impaired populations will reveal disorders affecting aspects of language which have Ibeen new!y described in modern and psycholinguistic work. This work has obvious diagnostic and therapeutic potential, as yet 1arGelyunexploited. The delineation and exploration of a domain of study restricted to grammar, in the technical sense, and the recovery of its levels and elements in language processing, has two implications for theory in neurolinguistics. The tirst is related to the domain-specificity of grammatical theory. We are encouraged to distinguish neural mechanisms underlying these aspects of language from those related to language in a broader sense. This distinction has, in fact, been implicit in neurolinguistics. With few exceptions, theories of language-brain relationships have not been directed at phenomena such as individual differences in verbal talents and situationally determined conditions for language use, but rather at what can be thought of as species-universal languagespecific representations and processing features of language, such as the locale (or lack thereof) of the representation of ‘sounds of words’, or the pathway (if any) whereby these representations are conveyed to ‘motor speech regions’. This traditional focus of neurolinguistic theories is quite consistent with the search for the neural basis for grammar and its processing. Modern linguistics and psycholinguistics essentially refine this approach, by presenting a detailed acount of the nature of the steps mediating sound and meaning. We are now in a better position to judge whether particular factors traditionally considered ‘linguistic’ in studies of aphasia, such JS word length (Howes, 1367), word frequency (Wepman et al., 1956), sentence probability (Goodglass and Kaplan, 1972) salience (Goodglass, 1968), etc., are part of the narrc! scope of linguistics and psycholinguistics--and hence newrolinguistics--or whether they reflect aspects of human cognition separa.te from those narrowly associateG1 with language per se, which interact with a ‘language faculty’. And we may look for neural mechanisms related to the host of elements, operations and levels which are identified by a grammar. The second implication of modern linguistics for neurolinguistic theory is related to the species-specificity of grammatical theory. Language capacities sre apparently restricted to humans; at least, there is no obvious mental system known in other species wit I-1comparable structural properties and communicative utility. It is entirely natural to ask what the material basis for this capacity is. We may consider the hypothesis that what is unique to the human species. with respect to language, is precisely those mental representations and as-
62
D. Caplan
so&ted psychological processing routines which constitute the its processing, The fact mentioned above, that the best accounts of lmguage structure and processing now available require elements and operations not found in theories of other mental capacities, suggests the legitimacy of this hypothesis. The implication for neurolinguistic theory, then, is that by approaching the neural basis for ‘grammar’ we may uncover a uniquely human aspect of neurobiology. The practical implication of both these hypotheses for research is that, to make progress along these lines,, we must separate those features of language which are intrinsic.to the system of grammar itself from tilose which are associated with its use but belong to other cognitive domains, and we must distinguish the effects of neurological disease upon the former from those upon the latter if we are to use clinico-pathological correlations as a means of investigating the neural foundations of language. I mentioned above that neurolinguistic theory can only relate theories of language and brain which are conceptually compatible. It is apparent that neural concepts employed in existing neurolinguistic theories are completely inadequate to the task of serving as the correlates of the richly detailed linguistic and psycholinguistic theories currently being articulated. The neural descrip tions used in neurolinguistic theories today are almost entirely directed to= wards the identification of macroscopically defined areas of the brain (gyri, parts thereof, and groups thereof)-structure9 which we cannot relate mechanistically to particular information-bearing units of language. Rectignizing the unreasonableness of expecting an answer in macroscopic neural terpms to a question such as “What neural structures differentiate the representation of Russian and English in speakers of these languages?” will permit us to appreciate rhe equivalent error in accepting as conceptually adequate the macroscopic neurological description of the material correlate of, e.g., a pazi of a grammar in a single language. k In sum, modern linguistics has at least three implications for theories of language-brain relationships. It suggests that we distinguish, at quite detailed levels of description, those features of language which are intrinsic to the mediation of sound and meaning and which constitute, ex hypothai, a natural domain ‘. f intellectu4 life, from those which belong to other cognitive capacities. It suggests that the delineaticjn of the former may constitute a characterization of a species-specific mental capacity, whose neural basis would then reflect uniquely human neural properties. Finally, it implies that we must go beyond the current level of neurological description to achieve a conceptually adequate correlation of neural and linguistic events. Progress has been made in all of these areas in the last ten years. Indeed, that these implications can form the basis for a research program into the
bospects for nkurolinguistictheory
63
aphasiasiswhat ~MC s from studies such asthose mentioned above. It seems
a reasonable hope that additional specific studies, perhaps facititated through the cleatian of inter-institution o~a~~zationat structures, will continue to deveflopaspects of neur&n uistic theory along these fines. for futua research programs invites comments on e potential of several approaches for the direct obsereal events duri language use (Desmedt, 1977; ress in neurolinguistic theory is sie of the performances of neurologiealillyimpaired po~ulati~ns~ and the correlation of abnormalities in performance with lesion site and type. Reasonable expectations along these lines seem to me to in&de the following: that we will greatly extend the range of uistic structures in which we detect abnormal performances; that we will begin the investigation of deviant Ilaguistic representations which are systematically related to normal representations; that we will more closely define the psycholinguistic processing of linguistic representations in individual patients through greater use of so-called ‘on-line’ measures of performance and greater awarncss of sources of variation in ‘off-line’ performance; that we will better characterize pathological performances as Gue to ‘deficits’, ‘additions’, and ‘modifications’ of normal processes (inriuding those processes normally present which generate errors in language use); that we will better appreciate how a failure at one level ramifies through the language system; and that we will extend studies of these aspects of language to patients other than those with recognizable aphasias due to stroke. The latter goal seemc to me to be part&lady important if we are to approach a neural theory of language in other than gross neuroanatomical terms, Correlation at’ particular abnormalities of language with lesion type- especially where different cellular lesions occupy similar grossly defined areas of the brain---may allow inferences regarding the cellular correlates of language through a Iogiieidentical to that which is now used for correlations between gross n:uroanatomical structures and aspects of language capacities. It would be foolhardy, to say the reast, to predict that: we will have an adequate nsurotinguistic analysis sf even one aphasic symptom or syndrome in ten or so years; certain!y it is unrealistic to expect a mechanistic neural analysis of any aspect of language, Work in the past ten years hascontributed a number of partial descriptions and the beginnings of functional explanations of symptoms and syndromes, in terms which begin to be compatible with existing linguistic and psycholinguistic work. In the process, it seems to me that a good deal of light has been shed upon what we would now want an adequate description and explanation of an aphasic performance to incl, le and what we would take to be the elements of an adequate neurolinguistic model of nor-
54
D. Caplan
mative abilities at both functional and neural 1elAs. Wopef~lly, work in the next-decade will continue to elaborate both the gorem1 and specific models that have begun to be developed within this framework. References Bradley, D. C., Garrett, M. F., and Zurif, E. (1980) Syntactic Deficits in Broca’s Aphasia. In D. CapIan (ed)., BiologiculStu$es ofMenrul Processes.Cambridge, Mass., MIT Press. .‘L; Chomsky, N. (1980) Rules und Represent&ions. New York, Columbia University Press. Coltheart, M., Patterson, K., and Marshall (eds.) (1980) Deep Dyslexia. England, Routledge. Dennis, M., and Whitaker H. A, (1976) Language Acquisition Followin& Hemidecortication. B&rim Lung., 2,472-482. Desmedt, J. E. (ed.) (1977) LanguageF lo ,‘.‘emisphericSpeciulizationin Mun: Cerebml l&w-Related Potentials.Basel, Karga. Fodor, J. D., and Frazier, L. (1980) Is the Human Parser an ATN? Cog., 8, 417-464. Garrett, M. (1978) Levels of Processing in Sentence Production. In B. Butteruc,.rh (cd.), La@~utr# Production Vol. 1. New York, Academic Press. Goodglass, H. (1968) Studies on the Grammar of Aphasics. In S. Rosenberg and J. H. Kapiin (cds.), Developmentsin Applied PsycholinguisticResearch. New York, MacMillan. Goodglass, H., and Kaplan, E.( 1972) Assessmentof Aphak and Related Disorders Philadelphia, Rnn., Lea and Febiger. Howes, I). (1967) Experimental Investigations of Language in Aphasia. In K. Saizinger and S. Salzir@er (eds.), Research in VerbalBehuviour. New York, Academic Press. Kean, M. L. (1980) Grammatical Representations and the Description of Processing. In D. Caplan (ed.), BiologicalStudiesof MentalRecesses. Cambridge, Mass., MIT Press. Klein, B. (1977) What is the Biology of Language? In E. Walker (ed.), Explorations in the fkdogy of Language. Montgomery, Vt., Bradford Books. Lecours, A. R., and Lhermitte, F. f 1969) Phonemic Paraphasias: Linguistic Structures and Tentative Hypotheses. Cortex, 5, 193-228. Morton, J. (1978) Word Recognition. In J. Morton and J. C. Marshail (eds.),PsJ&&t@fst&?s: srtuctures and Processes.Cambridge, Mass., MIT Press. Schnitzer, M. (1972) GenerativePhonology: Evidence from Aphasia. Penn. Swte Monagrapha, No 34. University Park, Penn., Penn. State University Press. Wepman, J. M., Bock, R. D., Jones, L. V., and Pelt, D. (1956) A Revision of the Concept of Anemia. J. Speech. Hear. Dis, 21,468-477. Wood, F. ted.1 (1980) Noninvasive Blood f:low Studios. Brain Lang, 9, I -148. Zaidel, E. (1978) Auditory Language Comprehension in the R@ht Hemisphere Followh~ Cumbral Commissurotomy and Hemisphercctomy, In A. Caramazza and E. Zurif (eds.), &un@u@ AC& sirionand LcrnguageBreukdown. Baltimore, Johns Hopkins Press. L.
Cognititm, 10 (1981) 65-70
65
@Elsevie; Sequoia S.A.. Lauwnne - Printed in The Netherlands
ialaiis a confoundednuisance,or: we be able to run any piycholinguistic experimentsat all in 19 ANNE CU’TLE fl* University of Sussex
Researchin e n;‘ive psychology tends to be paradigm-driven at the best of times, and the seven ies haven’t even been the beat of times. The most judicious attempts break the mould can be self-defeating; see Lockhart ( 1978), for example, bewailin the fact that the ‘levels of processing’ approach, devised by Craik an Lockhart (1972) its an attempt to inject more real-life relevance into memory research, was enthusiastically taken up by the field and developzd into a self-perpetuating paradigm. Psycholinguistics exemplifies the generai predicament. Its history over the past decade chronicles 2s much as anything else the continual discovery of new confounds. In order to facilitate this exercise, psycholinguists now conscientiously publish their materials in full. The more materials are published, the more confounds can be and are discovered. (Publish and perish.) In the following p:ges I will illustrate, by way of a few judicious examples, what this mears for the ordinary designer of psycholinguistic experiments; and of course, since I too wish to make an immortal contribution to the psycholinguistic literature, I may not refrain from pointing to a few confounds myself. Example 1: Wbt happened to fhe ambiguity effect in phoneme-monitoring
in the early seventies there was held to be an ‘ambiguity effect’ in yhonememonitoring; when the word preceding the tearget-bearing word was ambiguous, reaction times to detect the target were slower than when the preceding word was unambiguous (Foss 197Q), and this was true even when prior context made it quite clear which meaning of the ambigcc,us word was intended (Foss and Jenkins 1973; Cutler and Foss 1974). The effect was *Reprint requests should be sent to Anne Cutler, Laboratory of Experimental Psychology, University of Sussex, Brighton BNI 9QG, Em !and.
66
Anne Curler
explained as an increase in processing load due to the necessity to choose between the alternative meanings of the ambiguous word. But then along came Mehler, Segui and Carey (1978), and Newman and Dell (1978), who pointed out that the unambiguous control words in these experiments had more often than not been longer than the ambiguous words. Perhaps this added length had allowed the subjects a little extra processing time in the unambiguous condition. Indeed, when length waq controlled, the ambiguity effect disappeared; in fact by making ambiguous words longer than the unambiguous controls, it was possible to produce phoneme-monitorin,: rcacfion times which were faster following ambiguous than unambiguor+ words. Newman and Dell pointed out yet anotner confounding factor: th+ ambiguous words often began with sounds which were phonologica.lly similar to the target sound beginning the following word, whereas the unambiguous control words rarely did; judicious control of this factor also removed any indication of an ambiguity effect. The lesson to be learnt from this episode was, of course: take more care in constructing phoneme-monitoring experiments! No conclusions were drawn about the processing of ambiguous words. As it turned out? the hypothesis behind the early phoneme-monitoring work on lexical ambiguity was not entirely ill-conceived; it really does appear to be the case that occurrence of an ambiguous word in a sentence resuits in all possible meanings of the word being momentarily activated, irrespective of dlsambiguating context (Swinney 1979). Ht is not the case, however, that this produces processing difficulty measurable via phoneme-monitoring response time.
Example 2: Timing, frequency and intensity are very important’ As listeners process an utterance, they pay close attention to the prosodic (timing, frequency, intensity) variations; in tidct, they will1follow prosodic continuity at the expense of semantic continuity (Darwin 1975). Phonememonitoring response times are faster to targets on stressed words than to targets on unstressed words (Shields, McHugh and Martin 1974; Cutler and Foss 1977). Mloreover, listeners can use the prosodic contour to direct their attention to the most highly stressed parts of an utterance, leading to faster monitoring times (Cutler 1976; Cutler and Darwin 198 1). Obviously sentence prosody is a most important factor in sentervce comprehension, and t As the actress said to the bishop.
Making up materialsis a confbunded nuisance
67
ought to be taken into judicious consideration in designing and interpreting sentence processing experiments -particularly monitoring experiments. But is it? Well, usually not.*
Example 3: How to calculate word frequency All psycholinguists learn at their supervisor’s knee the importance of frequency of occurrence as a predictor of word recognition time; everybody controls for frequency. And so we should, since Whaley (1978) has showr. that it may well be the sirigle strongest influence on lexical decision reaction time. Theoretical explanations for the frequency effect in word recognition appeal to the structure and access ot the mental lexicon Folster 1976; Morton 1978. Among the things we know about the mental 1e::icon is that words regularly inflected, e.g., for fence and number, do not appear to be represented independently of their uninflected form (Gibson and Guinet 197 1; Murrell and Morton 1974; Stanners, Neiser, Hernon an:J Hall 1979). Therefore the frequency wiih which a particular lexical representation (say, pick) is accessed ought to be better approximated by the summed frequency of the uninflected with thH regularly inflected forms (i.e., pick + picked + picks + picking- 1S 1 in KuEera and Francis [ 19671 )-than by the surface frequency of’ pick alone (55)). Sure enough, the combincil frequency produces stronger frequency effects than the surface frequency (Rosenberg, Coyle and Porter 1966; Taft 1979i. It follows, then, that one does better to match experimental materials on tins combined frequency measure than on surface frequency. Some judicious Gxperimenters do this (e.g. Bradley 1978); most don’t . 3*4 Alas for us all, this is not even the whole story. Judicious matching 01. combined frequency, though more difficult than matching on surface *Consldcr for instance a recent word-monitoring experiment in which reaction times to the same target word wcte compared in three types of context: normal sentences, semantically anomalous but gramm:tically acceptable sentences, and scrambled strings of words. One of the important results of this cxrlcriment was that reaction times were fastcr in the normal s=.Wcncesthan in the abnormal, and faster ‘n the strings which were only semantically anomalous than in those which were syntactically anomai MIS as well. The speaker who recorded the experimental sentences was not aware which words were the targets. One might expect more prosodic cues to the location of content words (al! of the target words were content words) in the syntactically regular conditions than in the scrambled condition, and since sentence stress is semantically determined, more prosodic rues to sentence stress in the semantically normal condition than in the semantically anomalous conditioil. Thus these experimental materials may have confounded prosodic cues with syntactic and semantic cues. Notes 3 and 4. (Please see overleafl
68
Anne
Cutler
frequency, can be done. However, it turns out that we ought to match on surface frequency USwell. Taft (1979) found that combined frequency correctly predicted reaction time differences when surface frequency was controlled, but surface frequency also correctly predicted reaction time differences when combined frequency was controlled!
Consider a judicious psycholinguist constructing materials for an experiment comparing nou.ns, verbs and adjectives. Ideally he would like to create matched triples of an unambiguous noun with an unambiguous verb and an unambiguous adjective. They should be matched, as we have seen, on both surface and combined frequency. Naturally they should be matched on length. At this point it is already clear to the experimenter that the taak is probably impossible5 ; and he has not even begun to consider further variables on which they might be matched, such as association, age of acquisition.. autobiographical memory, categorizability, concreteness, digram frequency, imagery, goodness: letter frequency, number of meanings, orthographic regularity, meaningfulness, emotionality and recognition threshold Rubin 1980; Whaley 1978; Jastrzembski 1981. No easier task confronts the psycholinguist (designing a phclnememonitoring experiment in which the experimental sentences contain, say, words of high morphological complexity while the control sentences contain morphologically simple w0rd.s. On the basis of srveral models which describe specifically the process of monitoring for phonemes (Cutler and Norris 1979; Foss and Blank 1930), the expeimenter ca;l design the ,materials so that the 3A recent lexical decision experiment on prefixes, for instance, investigated words which could occur either alone or with prefixes. Words like pending, whibh have lower frequency than their pref”uredrelatives (e.g., impending) were compared with words like bark, which have hlg,ler frequency than prefixed forms (e.g., embark). Each word was matched on length and surface frequency with a non-prefixable control word; pending with picking, for instance, and bark with bull. It was predicted that there would be no reaction time difference between hrk words and their controls, but pending words would be responded to slower than their controls because the higher frequency pxefixed form would interfere; this pattern of results was :ideed fourid. But when one looks at combined frequency rhe matching turns out to be imbalanced. Thus while pendirrgand picking each have a surface tiequen‘v of 14, the combined frequency measures are respectMy 14 (i.e., no other form of peni occurs) and 151. In fact of the 20 patding words, 46 were less frequent than their controls or: the combined frqaency measure, whereas the lark words did not differ significantly from their controls in this zTp (II more frquent, 8 less frequent, one equal). The Tl~ornd&eAorge (1944) word count sums frequencies across regulslrinflections; unfortunately it is also 23 years more out of date than the Ku&-Francis (1967) count. ’ Even with the invaluabie assistance of Coltheart (1981).
Making up materials is a cmfounded
nuisame
69
monitoring response reflects the lexical characteristics of either the targetbearing word or the preceding word, according to choice. But the experimenter still has to match materials. The sentence prosody must be equivalent across experimental and control sentences, for example, the phonological similarity of target and initial sound of preceding word must be controlled. And most of the word recognition factors described above will be relevant. This is not nearly the end of the story. We have not yet begun, for instance: to assess the whole new range of possible confounds opened up by Marslen-Wilson’s observation. (Marslen-Wilson and Welsh 1978) that wlords differ markealy in the position of their recognition point, i.e., the point at which, counting from left t right, they become unique from all other words of the language. The prospects are glootny. If it goes on this way in the eighties, psycholinguists will literally b2 lost for words. Perhaps it is time for us to take matters-or rather, materials-into our own hands. Judicious choice of language in all our writings, for instance, combined with judicious extension of our fields of publication to other literary domains, could eventually allow us to exercise (judicious) control over the ratings assigned to words in future frequency counts. By way of a beginning, this essay represents a modest attempt to upgrade the frequency rating of the word judicious. !
References Bradley, D. C. (1978) CompututionalDistinctionof VocabularyType. Ph. D. Dissertation, Cambridge, Mass., MIT. Coltheart,M. cl9881 PsycholinguisticDatabase User Mar&Z. London, Medical Research Council Craik, F. I. Y. and Lockhart, 3. S. (1972) Levels of processing: A framework for mamory research. J. verb. Learn. verb. Beizav.,II, 671-684. Cutler, A. (1976) Phoner’e-monitcring reaction time as a function of preceding intonation contour. Percep. Psychophys., 20, 50-60. Cutler, A. and Da&n, C. J. (1981) Phoneme-monitoring reaction time and precedi.ng prosody: Effects of stop closure duration and of fundamental frequency. Percep. Pshychophys.. 29, 217-224. Cutler, A. and Foss,, D. J. (1974) Comprehension of ambiguous sentences: The locus of context effects. Pap1 presented to the Midwestern Psychological Association, Chicago. Cutler, A. and Foss, D. J. (1977) On the role of sentence stress in sentence processing. Lang. Sp., 20, l-10. Cutler, A. and Norris, D. G. (1979) Monitor& sentence comprehension. In W. E. Cooper and E. C. T. Walker (eds.), Senten& Rvcessing: Psycholir@stic Studies Presented to Merriii Garrett. Hillsdale, N J, Xrlbaum. Darwin, C. J. (1975) On the dynamic use of prosody in speech perception. In A. Cohen and S. Nooteboom (eds.), Structure und Rucess in Speech Pemeption. Heidelberg, Springer. Forster, K. I. (1976) Accessing the mental lexicon. In R. J. Wales and. E. C. T. Walker teds.), New Apprmches to ihngucgeMechanisms.Amsterdam, North-Holland.
70
Anne Cutler
Foss, D. J. (1970) Some effects of ambiguity npon sentence comprehension. J. verb. Leurn. verb. Betid., 9.699-706. FOSS,D. J. and Blank, M. (1980) Identifying the speech codes. Cog. Psychol., 12, 1-31. Foss* D. J. and Jenkins, C. Ma.(1973) Some effects of context on the comprehension of ambiguous sentences.J. verb. Learn. verb. Behav., 12. 577-589. Gibson, E. J. and Guinet, L. (1971) Perception of inflections in brief visual presentations of words. d v&. Lean. verb. Behcv.. IO, 182-189. Jastrzembski, J. E. (1931) Multiple meanings, number of related meanings, frequency of occurrence, and the lexicon. Cog. Psychol., 13, 278-305. K&m, H. and Francis, W. N. (1967) Computational Analysis of Present-Day American Eriglish. Providence, Ig 1, brown University Pres?, Lockhart, R. S. (1978) Method and content in the study of human memory. In J. P. Sutcliffe ted.), Concepnral Analysis and Method in Psychology: Essays in Horzour of W. M. O’Neil. Sydney, Sydney University Press. Marslen-Wilson,W. and Tyler, IL.K. (1980) The temporal structure of spoken language understanding. cog., 8, l-71. Marslen-Wilson, W. and Welsh., A. (1978) Processing interactions and lexical access during word recognition in continuous speech. Cog. Psychol., 10, 29-63. Mehler, P.., Segui. J. and Carey, P. W. (1978) ‘rails of words: Monitoring ambiguity. J. xrb. Learn. ,verb.Behav., 17, 29-35. Morton, J. (1978) Wcrd recognition. in J. Morton and J. C. Marshall teds.), Psycholinguistics Series 2: Structures and Processes. London, Elek Books. Murrell, G. A. and Morton, J. (1974) Word recognition and morphemic structure. J exper. Psychoi., 102.963-968. Newman, J. E. and Dell, G. S. (1978) The phonological nature of phoneme-lmcnitoring: A critique of some ambiguity studies. J. verb. Learn verb. Behav., 17, 359-374. Rosenberg, S., Coyle, P. 9. and Porter, W. L. (1966) Recall of adverbs as a function of the frequency of their adjective roots. J. Merb.Learn. verb. Behav.. 5, 75-76. Rubin, D. C. (1980) 51s properties of 125 words: A unit analysis of verbal behavior. J. verb. Learn. verb. Behav., 19, 736-755. StiG&, J. L., McHugh, A. and Martin, J. G. (1974) Reaction time to phoneme targets as a function of rhythmic cues in continuous speech. J. exper. Psychol.. 202, 250-255. Smers, R. F., Neiser, J. J., EIernon, W. P. and Hall, R. (1979) Memory representation for morphologically related words. J. verb,.Learn. verb. Behue., 18. 399-4 12. Swinney, D. A. (19?9) Lexical access during sentence comprehension; (Rz)consideration of context effects. Ir. verb. Learn. verb. Behav., i8, 545-569. Taft, Hi. (1979) Recognition of affixed words and the word frequency effect. Mem. Cog., 7, 263-272. Taft, M. and Forster, K. k. (1975) ILexicalstorage and retrieval of prefixed words. J. verb. Learn. verb. Behav., 14.638-647. Thorndike, E. L. and Large, I. (1944) The Teacher’s Word Book of 30,OaO Words. New York, Columbia University. Whaley, C. P. 0378) Wcrd-nonword classification time. J. verb. Learn. verb. Behcv,, 17, 143- 154.
Cognition, 10 (1981) 71-?8 @ Elsevier Sequoia S.A. !,ausanne - Printed in The Netherlands
71
Six tenets forevent JAMES E. CUTTING” Cornell University
We perceive events. Whether presented tachistoscopically in a dark room to a relatively passive: observer or presenteS over extended time in ,3n unbounded wilderness to an unfettered explorer, that which is perceived can usefully be described in terms of events. Events are our units of perception; indeed, they are our very units of exis,tence. Such a view is neither unique nor is it well-accepted. Many otht:rs take events central to perception, but perhaps at least as many do not. The reticence of the latter group is nlmost certainly due to nagging qurlstions of definition. What is an event anyway? What constitutes a nonevent? Sekcral classification schemes have been proposed (E. 9. Gibson, Referen#:;e note 1; J. 3. Gibson, 1979; Hej.der, 1959; Johansson, von Hofsten and Jansson, 1980; a.nd Pittenger and Shaw, 1975), but no consensual criteria exist for eventhood. I for one, however, am not bothered by this and am fully satisfied with the definition given by Webster: An event is “that which occupies a restricted portion of four-dime nsional space-time”, I like this definition, not only for its ingenuous accepta.nce of developments in twentieth-century physics, but for two other reasons as well. First, it implies that events have a structure that separates them from the rest of the world. Second, it explicitly mentions restrictions. It is the discovery and explication of those restrictions over the various classes of events that I fhd among the most exciting developments in perception at this time. Anything that czln be said ccllcerning restrictions on events is necessarily prematuia. Neverl[h,eless, there are more than a few ideas in this domain that have heuristic vAe. What ! propose here are six tenets for event perception that may serve as principles for discussion and discovery. For this presenta tion I have limited myself to unimodal, visual events. 1. Events have underlying structure One of the most fi-uitful ideas-in cognitive psychology has been that of underlying structure. A legacy from lin,i;uistics, the idea of underlying structures %upported by NIII Grant MII35530. Requests for reprints may be sent to J.E. Cutting, Dep.xWent of Psychology, Uris ik3, Cornell University, Ithaca, NY 14853, U. S. A.
72
.I. E. Cutting
b-as its great& appeal when one can find simplicity and harmony beneath a riot of ‘diverse surface strur ure forms. I believe such qualities can be demonstrated for event with wild!lg differing appearances(Cutting and Proffitj:, 198 1). One restriction on events, then, is that their surface structure must lawfully unpack in dimensions of space and time frow constraints dictated by their underlying forms. How this is done, of course, is not well worKed out, but the underlying structures are likely to contain two types of elements: Variants and invariants. The variants are inconstant across the event or differing events, whereas the invariants are constant. Although the variants are likely to give subtle richness to 8 wide variety of events, it is certain that the invariants will prove most useful for scientific study. 2. ‘?wo cbwes of invariants divide evwt structure: Topographic and dymmic ‘Event structure can divide many ways, but many people have parsed them in essentially the same manner. One of the earliest is due to KiihIer ( 1947, p. IO). From him we get notions of topographic invariantz. structural properties in space that hold over the course of the event, and dynamic invarkwts, ’ rules ,that govern the nat:~;cr:of change over the course of the event. Pittenger and Shaw (1975) parse events in essentially this same way. Perhaps the most compelling of the topographic invariants is the horizon ratio for determinin g the size of terrestrial objects (Gibson, 1979; Sedgwick, 1973). Looking at an outdoor scene the observer can easily determine the heights of various objects by the relation of their tops, bases, and the horizon. OR the proximal image all those things whose tops intersect the horizon, but pr@ct no higher, are as tall as the eye-height of the observer; all those that project above are taller; all those that fall below are shorter; and the relative ratios of base-tohorizon and base-to-top can be used. within a wide range to predict the S;Z of objects scaled to the size of the observer. A cornpIling instance of a dynamic invariant can be related to the same type of tiimulus scene. When the observer moves through the environment flow mrterns are created on the proximal image that radiate from the point tow&s which the observer is moving. The relative change in location of any set of points in the proximal image with respect to the fmed point invariably predicts the rellative&stance of that set from the observer (Gibson, Olum and Rosenblatt, 1955; Warren, 19’76). ‘1 am usiq the term &FTI&C in its uxnmonse~~ meaning, as it pertaim to motion or change, and irhdiug or e&uding forces that may be invoived. In tlds manner, my presentation is to be neutral with regard to the distiwion between kinematics and kinetics.
Six mcts for event perception
73
In sum, topographic invariants seem often to take the form of structural ratios in space for the various parts of an object in relaiion to om another; dynamic invariants often .take the form off similar structural ratios in time. These two types of invariants co-constrain one another-. Since psychology has spent most effort looling at topographic concerns in static images, the more novel work in event perception is in specifying dynamic relations in moving images. The most obvious way to do this, given the co-constraining nature of the invariants, is to use what we know about structural relations in space to educ,&e ourselves about structural relations in time-the dynamic li.nvariants. One such topographjc notion that is useful is that of wholes and parrs. 3. Dynamic .hvariants divide into those of wholes and those of parts. Events can have other events n,ested within them. Indeed, complex events are often composed of several layers of subevents. A hum&n being walking is one such event: The body moves through space, the arms move with respect to the body, and the lower arm moves with respect to the upper arm. If lights are mounted Ion the joints of walkers and the surround darkened (Johansson, 1973; Cutting, Proffitt and Kozlowski, 1978) familiarity cues of the display are removed, but the percept rerxins robust and the motions analyzable. The motion of the whole, or common motion (Johansson, 1950), for a walker is that of translation across the floor accompanied by a slight undulating motion. That is, the individual is seen to move horizontally with a small bounce. The parts, on the other hand, are not. seen to move separately across the floor. Instead, they are seen only as arms and legs swinging inquasi-pendular fashion wiih respect to the whole The point here is that the dynamic invariants specified for the whole need adifferent reference coordinate system than those for the parts: The appropriate coordinates 5x the whole are with respect to the environment, whereas the appropriate coordirlates for the parts are objectcentered. This is true not only for a human walker, but for other moving objects as well. Consider a rolling wheel. If 14_‘ltsare mounted on the wheel and the surround darkened (Duncker, 192); Proffitt,Cutting and Stier, 1979) one sees two types of motions. The common motion is that of translation across the surface and, if the lights are not mounted such that their centroid coincides with the center of the wheel, a hobbling motion as well. This is the dynamlc: invariant for the whole. The parts, on the other hand, rotate in circles about the2 centroid, and rotation is th.e dynamic invariant for the ptis. Agab, environmental coordinates are most useful for the motion of the whole;, and object-centered coordinates for the motions of the pt;rts.
74
J. E. Cutting
A question remains, however, 3s to how these two types of dynamic invariants are perceptually segregated. The scheme that appears most fruitful is connected to an old Gestalt idea: 4. Dynamic hariants divide according to a minimum principb Kijhler (1920) suggested that physical Gestalten tend towards states requiring minimum energy for their mairntenance. They are simpler in terms of the constraints of natural laws. Borrowed from physics, this idea has seen periodic popularity in perception, and I hope it is again on the rise. Restle (1979), for example, presents 6; I’ramework in which simplicity is a guiding principle for motion perceptian. Simplicity has always presented thorny problems to both psychologists and philosophers (Sober, ‘r975), but in the domain of event? we might profitably look towards simplicity as minimal descriptions of change given the universe of possible descriptrolls. The perceptual system might well choose the dynamic invariants for wholes and parts on grounds of minimality. Consider again a rolling wheel in a dark room, this time with two lights mounted on the perimeter 90” apart(MSrjesson and von Hofsten, 1975 ; Proffitt and Cutting, 198Oa). What is seen is a wheel-like object hobbling across the field of view, following the pat.h of a prolate cycloid. The two lights form a group, that is they instill figural coherence, but they do not look like a smoothly rolling wheel. A msmum principle might apply to the perception of this event in the following manner: The two lights are seen to be, not 90” apart, but 180”’ apart on a relatively smaller object rotating about a midpoint (centroid) between them. In this manner two interrelated factors are minimized : a movement factor where the momentary relative vectors sum to zero, and a spatial factor where the squared Jengths of the moment arms from their center of rotation sum to their smalfest value. The residual from this operation is the common motion of the whqle-the prolate cycloidal path. [It is clear that thi 1 procedure does uot work in all cases, but it doea,in the vast majority (Proffitt and Cutting, 198Oa, 198Oa). Moreirver, where it does not another minimum principle appears to operate on common motions.] h this manner, then, the perceptual system can determine the dynamic invariants of the parts through application of a minimum principle, isolate that factor, then determine the dynamic invariants of the whole. In each step of this process the topographic invariant, the relative spatial separation of the G&s, guides extraction of dynamic invariants. Logically speaking, the reverse procedure would as10 be true but since the topography of this system is relative%ysimple and the dynamics relatively complex it seems prudent to consider
Six tenets fqi- event percepthri
75
only the procedure outlined above. What I have not described yet, however, is a joint produci of the dynamic invariant of the parts and the topographic invariant, and what that product is good for. 5. Dynamic and topographic invariants yield a center of moment Most moving or changing parts of a coherent object have systematic reference to a single point .within that object. That is, almost all points can be said to mow around 1 point that serves as the &in in the coordinate reference system of the object. If the object is a fully visible rolling wheel, all points in the wheel rotate around the center of the axle; if the object is a pendulum, all points in the bob and pendulum arm oscillate around the pivot; and if the object is a lever, all points on both side of the lever ?cm move around the fulcrum. The general scheme can proceed for all rigid objects with pliable joints (such as the human body), and many elastic objects, particularly those where the elasticity is not uniform. My colleagues and I call this point the center of moment (Cutting et al., 1978). Often there is one cent&r of moment in an event, such as the fixed point in a flowfield discussed under Tenet 2. But perb.aps more often there is a whole hierarchy or system of centers of moment. Consider again the human body when walking: The lower arm moves around its center of moment, :hc elbow; the whole arm moves about its center of moment, the shoulder; and the whole upper body (in fact the whole torso) moves &round its center of moment, which lies near the waist and is determined by the relative widths of shoulders and hips (Cutting et al., 1978). Thus, vith the human body we have subevents embedded within events (forearm movement within. the armswing within the step cycle) and lower-order centers of moment superseded by higher-order centers of moment (the elbow by the shoulder, and the shoulder by the center of moment of the torso). In this manner, events can kave more than one center of moment, but they ordinarily are not arrayed in a structure other t?ran a pure hierarchy with a single center at the topmost node.2 2By now one should have detected a broken symmetry in my discussion. Dynamic and topographic co-constrain one another, and their product is a center of moment, a point in space around which all changes occur in time. Since most events have beginnings, middles, and ends, it seems luite reasonable, by analogy, to argue that there should be a point in time around which all changes occur in space. After all, events should hauretemporal centers as well as spatial ones. Like the rest of event perception, there seems to be little work on this concept (but see Fowler, 1979; Morton, Marcusand Frankish, 1976). By inverse analogy to center of moments I am tempted to call these “moments of center”, invariants
76
J. E. fitting
What I have said thus far, particularly with regard to centers of moment, is hardly the stuff of which psychology is normally made. It is more akin to something like ecological mechanics. What makes it psychology,in my opinion, is my last tenet:
6. Centers of moment are perceptually useful Centers of moment, as products of “he dynamic and topographic invariants, can be used by perceiver to make decisions about what they are viewing: The location of the highest-order center of moment in a human walker, for example, m~j; be used to determine gender (Cutting et al., 1978); the location of the center of moment (centroid) of a configuration of lights mounted on a rolling wheel may be used to determine how wheel-like the movement of that configuration appears (Proffitt eta& 1979; Proffitt and Cutting, 198Oa, 19806); the location of the centerof moment of the profile of an aging human head may be used to determine best examples of the aging process (Pittenger and Shaw, 1975; Cutting, 19783); the location of the center of moment in a flowfield may be ausedto determine the direction in *whichan observer is going (or where she has been)(GibsonetaZ., 1955;CuttingandProffitt, 1981);and thelocation of the center of moment of the rotating night sky (Polaris) can be used by migratory song birds to determine the direction of migratory flight (Emlen, 1975; Cutting and Profftt, 1981). In these and perhaps many other different types of events, one aspect of underlying structure is shared--a center of moment-and in all cases this point is perceptually useful. I find it compelling that such umty and harmony can be found under the farragoof suchdiffereIt surfam forms.
Of clouds and docks Karl Popper (1972, p. 207), in an essay on the problem of rationality and the freedom of man, broached the issue of clouds versus clocks as prototypes for structures of physic3 systems: My clouds are intended to represent physicalsystems,which like gases,are highly regular,disorderly,andmwe or kss unpredictable. I shall assumethat we havebefore us a schemaor arrangement in which a verydisturbedor disorderlycloudis placed on the left. On the other extremeof our arrangement, on its right,we may placea very reliablependulumdock, intendedto representphysicalsystemswhichareregulaz,orderly,andhighlypredictablein theirbe‘havior.
Six tenets jbr event percep;ion
77
If one substitutes for the notion of a physical system the notion of an event, then one can imagine an array of events, proceeding from irregular and disorderly to reg Oiarand orderly, from clouds to clocks. One might suspect from my description of human walkers, rolling wheels, aging faces, expanding flowfields, and rotating night skies, that all those things that we are wont to call events are very clocklike. Clearly, this is not the case, A moving cloud is a particularly good counterexample. As it scuds across the sky i.t changes shape, and these changes do not occur about any particular point within or near the cloud. Clouds swirl and billow unpredictably and take on new contours without necessarily undergoing coherent and -qecifiable mathematical trans,formzLtions. In essence, a cloud has an indeterminably large number of variants and few if any invariants, yet a moving cloud might easily be classified as an event. It sezms possible that among all possi.bie events there may be many that are more cloud-like than clock-like. These will probably not be amenable to the type of analysis given above. Nevertheless, I do not believe it crucial that the analysis pmsented here necessarily applies to all events. I claim only that this analysis, looking at various aspects of underlying structures, is useful in a wide variety of pe:rceptual situations. As Wigner (1967, p. 42) noted: We have ceased to expect from physics an explanation of all events, even in the gross structure of the universe, and we aim only at the discovery of the laws of nature, that is, the regularities of events. The same should be true of psychology.
References Biirjesson, E., and voc Hofsten, C. (1975) A vector model for perceived object rotation and translation in space. Psycho! Rer., 38, 209-230. Cutting, J. E. (197&r) Gemration of synthetic male and female walkers through manipulation of a biomechanical invariant. Percept., ?,393-405. Cutting, J. E. (19786) Perceiving the geometry of age in a human face. Pwcep. Psychophys., 24, 566-568. Cutting, J. E. and Proffitt, D. R. (1981) Gait perception asan example of how, we may perceive events. In R. Walk,and H. L. Pick, Jr., (eds.), Intersensory perception and sensorsr integration, New York, Plenum. Cutting, J. E., ProffLt, D. R., nnd Kozlowski, L. T. (1978) A biomechanical invariant for gait percey tion.J. exper. Psychol. -Am. Percep. Perf, 4, 357-372. Duncker, K. (1937) Induced motion. In W. D. Ellis (ed),A Sourcebook of Gestdt Psychology. London: Routledge & Kegan Paul, (Originally published in German, 1929.) Emlen, S. T. (1975) The stellar+rientation system of a migratory bird. Sci. Aliser., 233. 6% 102-l 11. Fowler, C. A. (1979) “Perceptual centers*’ in speech production and percepti(Jn. Percep. Psyc hophys., 25, 375-388.
78
J. E. titting
Gibson, J. J. (1979) The ec&gica~ approach to visual perception. Boston, Houghton Mifflin. Gibson. 1. J., Ohm, P., and Rosenblatt, F. (1955) Parallax and perspective during aircraft landings. Amer. J. Psychol, 67, 372-385. F, (1959) On perception, event structure, and psychological environment. Psychol. Is., I (3) I-123. Ishansson. G, (1950) C?mjiirations in event perception. Uppsala, Sweden, Almqvist 8~ Wiksell. Johansson, G. 41973) Visual perception of biological motion and a model for its analysis. Percep. f$VCh@lJ%, 14, 26i Jahanswn, G., von Hofsten, C., and Jansson, G. (13803 Event perception. An. Rev. Psycho& 31, 27-66. Kiihkr, W. $1937) PhysicalGestalten. in W.D. Ellis (ed), A sourcebook of Gestalt Psychology. London, . Routledge & Kegan Paul, (Originally published in Germaq 1920). Ki4h%r. W. (1947) Gestalt psychdogy New York, Liveright. Morton, J., Marcus, S., and Frankish, C. (1976) Perceptual centers (P-centers). Psychol. Rev., 83, 405-408. Wenger. I. P. snd Shaw, R. E. (1975) Aging faces as v&al-elastic events: Implications for a theory of nonrigid shape percepti0n.J. exper Psychoi. Hum Percep. Perf.. 1. 374--382. Popper. K. R. (1972) Objectiue knowledge. New York, Oxfore University Press. ~ekier,
-211.
Pruffi& D. e.. and Cutting, J. E. (198012)An invariant for wheel-generated mct5ns and the logic of its determination. Pmcep.. 3.435-449. Proffitt, D. R. and Cutthlg, J. E. (198Ob) Perceiving the rnntroid of curvilinearly bounded rolling s &a~es. Percep. Aychophys., 28 484-487. fkoftitt, D. R., Cutting, J. E., and Stier, D. M. (1979) Perception of wheel-generated motions. J. expcr. p&w~.
Hum Percep. Perf.. 5. 289-302.
Restle, F. (1979) Coding theory of the perception of motion configurations. Psychol. Rev., 86, l-24. %&wick, H. (1973) The visible horizon: A pote.ntkl source of visual information for tlie perception of size and distance. PhD Dissertation, Cornell University, Dissertation Abstracts Internutional, 1973i74,346,1301-1302. (University Microfilms No. 73-22,530). Sober. E. (1975) SimpIicity. New York, Oxford University Press. Wmen, R. (19761 The perception of egomotion. /. exper. Psychol. Hum. Percep. PerJ, 2.448-456. W@x. E. P. (1967) Syetries artd reflections. Bloomington, Indiana: Indiana University Press. Reference Note i. Gibson, E. J. A C%ssifiiation of events for the study of event perception. Paper for symposium on Even? Perception, American Psychologi-d Association, Chicago, September 1975.
Cognition, 10 (1981) 791-84 0 Elsevier Sequoia S.A., Lausanne - Printed in The Net herlands
79
Infants, speech,and language:A look at some connmtionls~ FtTER
D. EIMAS*
Brown UraivetGty
A major goal of those concerned with the cognitive developmen’l of the humarl species is to explicate the nature of the initial state of the infant-to describe, in essence, those biological endowments that accompany :he infant into the world and support cognitive growth. Of course, they also seek to understand the manner in which this inheritance int::racts with the infant”s continuous experience with. the environment to shape the processes by which knowledge is acquired as well as to shape the very form of knowledge itselff. The study of speech p-crceptinn in prelinguistic infants is no exception. Our motivation for this endeavor was and continues to be to provide a description of the initial abilities to process speech. We seek, as well, the course of development of these abilities, which is to say, the moduiations that occur with experience, and we seek understanding as to how these abilities, together with other faculties, both cognitive and linguistic, serve anal ultimately permit the acquisition of language. For some,‘a goal of this research has also been to construct a processing model r”orthe perception of speech at ihe level of phonetics, a goal often arisir-: from tile belief that by studying the origins of speech perception whereit is uninfluenced by higher-order linguist’c knowledge, we might be provided with special insights into its nature. It is of course difficult to evaluate objectively the progress that has been made toward achieving these goals, especially when one is &ill U!y occupield with the complexities and awesome sophistication of the infant’s ability to perceive speech. But the difficulties of evaluation pale before the difficulties of predicting ‘.. . the main outcomes and theoretical breakthroughs in the next decade’ (.I. Mehler, personal communication, 1980). Worse yet is the thought that our editor, J. Mehler, will ask us to comment on these evaluations and predictions a decade from now. z Evaluation of any research field must ultimately rest, not on the goals, but on the knowledge that has been gained. In the past decade, a number of issues have been addr Tssed and at h:ast partial answers have been forthcoming. Consider, for exampie, the issue of categorical perception. One of the more re*The preparation of this discussion was supported by grant MD0533 l-1 1 from the National lnstitute of Child Health and Human Development. I thank Joanne L. hfiller for her comments and criticisms of an earlier version of this paper. Reprint requests should be sent to: Pr Zer D. Eimas, WalterS. Hunter Laboratory of Psychology, Brown University, Providence, Rhode Bland 02912, U. S. A.
$0
Peter LX Eimas
liable findings in the field of speech perception is that mature listeners under a variety of experimental procedures and surely during the course of cornprehending the meaningfulutterancesof natural conversations,perceive speech categorically. We are aware of the differences between .the phonemic categories of our language, for example, between voiced and voiceless consonants, between stops and semivowels, and so forth. L-Iowever, we are tdrely aware of the acoustic variation that occurs naturally within any sigle category along certain critical acoustic continua, voice onset time, for example. At issue was whether the process of categorizing the natural world of speech, the mapping of acoustic variability onto discrete phonemic categories, was a result of experience, perhaps the type of experience that occurs when the child first be,$ns to notice that some acoustic distinctions signal meaningful linguistic contrasts, .whereas others do not. The issue had a simple solution: experience was not necessallj for categorization at this level of perception. Infantb from the age of one month were found to be capable of categorizing speech sounds, as evidenced by marked discontinuities in their discriminability functions. This was true for consonantal distinctions based on the acoustic information for voicing, place of articulation, and, in several instances, manner of ar@cAation, as well as when the information signalled a contrast between very brief vowels (for reviews of these findings see Eimas and Tartter, 1979; Jusczyk, 1981). Moreover, the ability of infants to categorize acoustic information for speech is not simply a matter of mapping the values along a single acoustic continuum onto discrete categories, but rather involves, at least at times, consideration of the information along other dimensions as well. Thus, for example, Eimas and Miller (1980) have found that in two- to fourmonth-old infants the discriminability of a difference in the duration of fcs;mant transiftons, a cue for the manner distinction betwee stops and semivowels in adult listeners, depended not only on the particular values of transition duration but also on the durati.on of the entire syllable. Virtually identical effects were obtained earlier by Miller and Liberman ( 1979) with adult listeners. It would seem then that the perceptual.system of infants is well suittiLl to begin the process of forming categories that will eventuate in the phonemic categories of one’s language. Moreover, the nature of the categorization pro, cess appears to be little altered by experience. This is nor to claim that experience has no influence on the perception of speech or on its production. Obviously it does. We not all speak a single language corrposed of common phonemic categories. nor do we find it equally easy to recognize all possible phonemic differences;those that are a part of our own language are often more readily detected than contrasts that are absent from our language. One aspect of infant speech research has been to investigate the role of experience on early perception by comparing the relative
Infants, speech, and language: A look at some connections
81
discriminabihty of acoustic information for phenemic contrasts that are or are not a part of the parental language. However, the task is not as simple as was originally believed. MacKain (1980) has noted that the assumption that a language whi.ch does not have a particular phonemic contrast provides a different acoustic environment from a language which does have this contrast, may rest on prior assumptions concerning the manner in which the infant segments the speech signal. This is especially the case for voicing distinctions based on prevoicing, the phonetic feature most often examined for possible influences of experience, In addition, of course, the presumed differences in the acoustic environments of infants must be verified empirically, something which has not yet been done. Thus, whrle it may be in fact quite difficult to assess experiential effects on speech perception at these very early ages, the research in this domain has by no means been in vain, What has emerged is the rather clear conclusion that infants from a variety of linguistic environments are capable of distinguishing many, and perhaps all (although this would be disputed by some), of the acoustic distinctions that underlie adult phonemic contrasts. There is also a growing research effort concerned with the infant’s abilil-y to form equiv?!ence classes based on the sounds of speech. The ability to form such classes is at least a rudimentary requirement for perceptual consta.ncy. This research began with a study by Fodor, Garrett, and Brill(l975) and has been continued by Kuhl (e.g., 1980) who has found quite convincing evidence that infants by the age of six months are able to form equivalence classes based on the vowel categories [a], [i] p and [ 41. She has also found that equivalence classes can be based on consonantal categories in similarly aged infants. In summary, the research of the past decade has establisued that young infants are capable of categorizing the sounds, of speech as evideulced by discontinuous discriminability functions and the fc;mation of equivalence categories. Given the generality of these abilities alld the early age of their appenance, their determination is reasonably ascribed to biological influences that are capable of modification by experience with the parental language. The significance of these sophisticated processing abilities and their genetic determination are not difficult to discern. For example, without the ability to form equivalence classes, the task of mapping the sounds of speech onto referents becomes one of ;tcquiring an infinite number of associations-clearly an impossibility. Perception in a categorical manner permits the encoding of speech in a digital form, a cognitive economy that should permit the greater allocation of resources to the absttaction of linguistic rules than if it were necessary to store speech in an analogue form. While an absence of an efficient temporary code might not render the acquisition of language impossible, it would certainly result in slower, less successful learning. Finally, if these pro-
82
Peter 19. Eimas
cessing abilities were not a part of our biology, it is difficult to imagine how an infant could learn or be taught to process speech in ways that will come to have significance for language and its acquisition, but that at the time of learning seem n-relevant to the communicative abilities of the infant. With regard to the development of a model err speech perception, progress has teen more limited. Descriptions of the perceptual process have tended to be based on a knowledge of articulator-y principles (e.g., Liberman, Cooper, Shankweiler, and Studdert-Kennedy, 1967; Stevens and House, 1972) or on analytic procedures based on feature detectors (Abbs and !Sussman, 197 1; Eimas and Corbit, 1973; Stevens, 1975). The former set of models encounters difficulties in that they do not readily permit explicit and testable predictions and they require assumptions of considerable innate knowlledge if they are to accommodate the data from infants. Feature detector models, on the other harrd, while capable of explaining much of the infant data and of yielding testable hypotheses, require extreme modifications if they are to account for the mass of empirical findings that are often contradictory to the metatheory of detectci models of perception. In the first place, a number of studies have indicated a need for multiply-tuned detectors, which, because of the number of detectors involved, threatens to topple a potentially parsimonious model of perception as a result of a surfeit of theoretical mechanisms (Eimas and Miller, 1978). Second, and more damaging to the detector model, are recent studies (e.g., Diehl, Elman, and McCusker, 1978) that indicate that the manner in which selective adaptation exerts its influence, by biasing operations and not by sensory fatigue, is not consistent with detector theories. How ‘we will reconcile t::ise latter findings with the considerable evidence, that is in accord with detector theories of perception, is not apparent at present. Based on present trends, it will probably be the case that the task of constructing explicit models of speech processing will not occupy many in the field of -peech perception or infant speech perception. Perhaps this is just as well. Extensive new kno<wledgein a number of domains is obviously required before successful theoretical descriptions can be formulated. For example, more precise definitions of the acoustic information that signals phonemic contrasts are needed. At the heart of this issue is, of course, the apparent lack of invariance m the speech signal. Although the search for invariance continues (e.g., Stevens and Blurnstein, 1981), success still appears to be considerably less than complete. Moreover, should these lines of investigation not find the sought-for invariance and should the discovery of additional contextual effects or trading relations among the acoustic cues for speech continue, models of speech perc;ption will by necessity require a level of complexity that far exceeds that which has been assumed up to present times.
Injbts, speech, and language: A look at some connect&m
83
In addition to finer stil~lus definitions, further investigations, with more sensitive methodologi.es, are needed if we are to be able to describe the kvels of processing required to derive the initial representation of speech. Tlicce is ample evidence that there are multiple levels of analysis, but exactly how many and the nature of these processing procedures remain matters for dzbate. Should progrless be forthcoming on these issues, theory development may well become a more central occupation of the field of speech perceptionan endeavor that will be welcome by many as a sign of past progress and future promise. Infant speech resea.rch, per se, will continue to provide descriptions of the initial state. As experimental procedures become more sensitive and more adaptable to the demands of older infants, we should be in a better position to investigate the effects of experience on the perception of speech and the manner in which this form of competence serves the acquisition of full linguistic competence. Further delineation of the biological constraints on the perception of speech, and ipso facto on language and its acquisition, should help to uncover the nature of these constraints, that is, whether linguistic or cognitive or perhaps both, as well as help to describe the course of evolution of human language. I would argue. that data relevant to these issues, while always difticult to obtain, will be more readily forthcoming from investigations of infant speech perception than from investigations of language learning at higher levels of abstraction. The reasons for this are that input-output relations, stimulus definitions, and definitions of mature levels of processing are more available to those who investigate the former than the latter. It is these issues that will proviC;; the motivation for the study of infant speech percep tion as well as shape specific research in the decade to come--unless, as is always the case in science, new discoveries in the 1380’s alter our conceptions of what we seek to understand. References Ebbs, J. H., and Sussman, lf. M. (1971) Neurophysiological feature detectors and speech perception: A discussion of theoretical implications. J. sp. learn. Res., 14. 23-36. Diehl, R. L., Elman, J. L., and McCusker, S. B. (1978) Contrast effects on stop consonant identil‘icati0n.J. exper. Psychoi: Hum. Percep. Perfi. 4, 599-609. Eimas, P. D., and Corbit, J. D. (1973) Selective adaptation of linguistic feature detectors. Cog. I’su_ chol.. 4,99-109.
Llmas, P. D., and Miller, J. :_., (1978) Effnr!; of selective adaptation on the perception of speech and visual patterns: Evidence for fe:ature detectors. In R. D. Walk and H. L. Pick, Jr., (eds.), Percep tion and experience. New York, Plenum Press. Eimas, P. D., and Miller, J. L. (1980) Contextual effects iu infant speech perception. Sci., 209, 11401141.
84
Peter D. Eimas
Eimas, P. D., and Tartter, V. C. (1979) On the development of speech perception: Mechanisms and analogies., In H. W. Reese and L. P. Lip&t (eds.), Advancesin childdeveiopment and behavior. Vol. 13, New York, Academic Press. ‘Fodor, .J. .4., Garrett, M. F., and Brill, S. L. (1975) Pi Ka Pu: The perception of speech sounds by prelinguistic infants. fercep. Psychophys.,18, 74-78. Jusczyk., P. W.(1981) Infant speech perception. In P. D. Eimas and J. L. Miller (cds.), Perspectiveson rhe study ofspeech. Hillsdale, NJ, Ermaum Associates. Kuhl, P. K. (1980) Perceptual constancy for speech-sound categories in ear!y infancy. In C. H. YeniKomshian, J. F. Kavanagh, and C. A. Ferguson (cds.), Child phonology. VoL 2: Perception. New York, Academic Press. Libernran, A. M., Cooper, F. S., Shankweiler, D. S., and Studdcrt-Kennedy, M. (1967) Perception of the speech code. Psychol. Rev., 74-4 3 l-461. MacKain, K. S. (1980) Qn Assessing tne Role of Experience on Infant Speech Discrimination. Unpublished manuscript. Miller, J. L., and Liberman, A. M. (1979) Some effects of later-occurring information on the perception of stop consonant and semivowel. Percep. Psychophys., 25,457-465. Stevens, K. N. (197 7) The potentiai role of property detectors in the perception of consonants. In G. Fant and A. A. Tatham f cds.), Auditory analysisand the perreptior of speech. New York, Academic Press. Stevens, K. N., and Blumstein, S. E. (1981) The search for invariant acoustic correlates of phonetic features. In P. E, Eimas and J, L. Miller (eds.), Perspectiveson the study of speech. Hillsdale, NJ, Erlbaum Associates. Stevens, K. N., and House, A. S. (1972) Speech perception. In J. Tobias (cd.), Foundationsof moderr: auditory theory. Vol. 2 I New York. Academic Press.
C.bg&bt?, 10 (1981) 85--88 @ Ekevim SequoiaS.A., Lausanne- Printedin The Netherlands
85
ic functionsand mentd rpam GiLLES FAUCONNtER* h~;S
A p~pub, although not always explicit, view of the relationship between psy choli a,uistics and hnguistics, as scientific fields, used to run something like this ,inguists, using the powers of pure reason, plus a good, but revisable, theoretical framework and a decent amount of data in the form of grammaticality judgments, set out to discover fhe structure of language (i.e., levels, rules, conditions on rules and levels...). Psychologists are provided with the linguists’ results and their task is to find out experimentally how such structures are processed and also, thereby, to give the postulated structures additional support. But the rediscovery of some epistemological truisms has perturbed this simple scheme. First, the structure attributed to the output of a system is not necessarily reflected within that system (the so-called E:st order isomorphism fallacy, cf., termites, planimeters...). Second, data is not theory-independent: a theoretical framework not only ‘explains’ data, but specifies what kind of data is actually relevant, legitimate..., and how it is to be gathered. Third, the concepts needed for classifying and theorizing do not arise magically out of combinatorial analysis: tT,ey are either invented, which requires not only formal and empirical thoroughness but considerable imagination as well, and/or borrowed from other domains. My own work, during the last few years, has focussed on issues where such epistemological problems arise in relation to natural language logic, pragmatics, speech acts and syntax. These studies htive triggered extensions of the range of relevant data, sometimes within language (scalar phenomena, implication reversal, quantification across discourse), sometimes beyond (social rituals, anticipation, principles of intemuption), The corresponding shift in theoretical emphasis is, characteristically, away from the formal representation of ‘under’ linguistic structure to the explicit identification and explanation of the processes which mentally set up discourse on the basis of various factors. Discourse, under this view, is in no way a sequence of sentences or propositions, but rather a separate mental construction triggered by sentences, context, assumptions, et c,, and performed by a speaker, or by a listener. Communication, as opposed to disburse, involves partial matching and negotiation of these mental constructions. *Reptint requestsshould be sent to Gil& Fauconnier,9, Rue des Guillemites,75004 Paris,France.
86
GillesF~iucmnier
This view is not assumed a priori. On the contrary it emerges from the failure of contemporary ‘structural’ approaches to deal with basic semantic phenomena. Nor, of course, is it restricted to my own work; it is part of what I see as a very general trend in linguistics, psychology, and other social sc%nccs. A prominent portion of the relevant work, of possible interest to psychologists, is the elaboration of the ‘mental space’ framework, which makes explicit claims, on largely linguistic grounds, regarding the mental constructions involved, how they work, and the range of data they account for. Language, perhaps not so surprisingly, displays remarkable logical intricacies which are not (directly) attributable to its overt syntactic structure, transparence/opacity, quantification., muItiple scope ambiguities, control of variables, projection of’ presuppositions, positive and negative polarity... A frequent theoretical ap preach to such problems has been to postulate hidden representations or ‘logical’ forms of sentences which might directly mirror the observed properties. But, by and large, such approaches have not revealed the general principles actually at work, and do little more than code ad hocl” the properties of particular fragments. Recent work on pragmatic funstions, correspondences., and images, offers a conceptually different, theorelically more promising, and empi:ically broader, system for understa,gding natural language logic. The philosophy behind the theory is that logical complexity in natural language does not follow from corresponding structural complexity of sentences, but rather from the complexity of processing configurations that simple syntactic structures ‘permit, and in other cases merely from the complexity of logic itself as a (partially inadequate) instrument of measure. Accordingly, the load of explanation is shifted from sentence structure to discourse processing. More precisely, it is shown how sentences set up ‘mental spaces’ (images, beliefs, desires, hypotheticals, time slices, fictions, umwelts) linked to each 0the.r by pragmatic functions similar to the ones that operate in ordinary reference. The fact that a given sentence may be compatible with several space configu.rations, and that inference’operates within, not across, spaces, explains why superficial *logical’ paradoxes arise, such as failure of substitution, invalid inferences, presupposition transfer, or lack of it. The approach and the corresflonding framework are defended on internal linguistic grounds, but the necessary view that processing is involved rather than structure forces a psychological interpretation: mental spaces, pragmatic mflespondences, flowing etc. are theoretical constructs which must be retfleeted psychologically in order for the approach to be consistent. Among the results obtained using this approach, let me single out the following: --It turns out that superficiallytricky and apparentlydifferent %emanfic’properties of sentences such as opacity/transparency, specific/non-specificcontrasts, time
Pragmaticfunctions and mental spires
87
ambiguities in deiinite descriptions, modes of reference to non-existing objects, follow straightforwardly from a single ‘Princ.ple of Identification’ across spaces, which itself follows trivially from the definin; property of pragmatic reference functions in general. The phenomena, then, do n:t really reflect properties of sentences, but rather properties of the space configu:ations compatible with particular sentence sfrucfurcs, given the correspondence principles, -The same holds for many other cases of perceived ambiguity: because of the correspondence principles (such as: if a (of space E) corresponds to b (of s,paceI)(pragmatical&, in the mental construction of discourse), a description of a may be used to identify b) the same sentence may be used in connection with selleral (sometimes an infmhcy of) quite different space configurations. This in no way entails that the senten= itself has several structures. Positing such structures would amount tti useless and arbitrpty reduplificatiolb, but, even worse, it would in fact be impossible, given the cases where the numbrrr of relevant interpretations increases unlimitedly with the number of spaces in ttie discourse configuration. Not surprisingly, then, sentences turn out to be considerably more ‘ambiguous’ logically than is commonly perceived. For example ‘In that picture, Mary is t iller than Harriet’ can be shown to have at least twelve logically distinct interpretations, and ‘Max thinks that the president of the Republic is Llore corrupt than the mayor of Paris’ has at least 192! It would, I think, be unpleasant to have 192 logical forms corresponding to that sente:nce. By contrast, the possibili:y of 192 space-correspondence configurations in harmony with the sentence is a simple property of the functions at work. --New and non-trivial properties of presuppositions, flowing and transfer, are revealed. They too are conditioned by space correspondences, and depend on systems of belief and social settings. The old ‘projection’ problem turns out to be an artifact of logically or syntactically based,approaches and disappears. -.Many specifically grammatical pro -erties, such as sequen4.e of tense constraints or ,indicative/subjunctive contrasts in relative clauses, are linked to space configurations. --:Reliminary investigation suggests that the framework appiies very naturally to language in non-oral modes, such as ASL.
spaces’ constitute one area for which an epistemological shift in theory and methodalogy seems warranted. In different, but conceptually related work, I have studied speech act phenomena in terms of the wider notions of social ritual a-id logic of action, arguing that independently attested principles pertaining to the transformation of rituals in general and the equivalence of some interruptcitd patterns of action to completed ones provide the key to speech act phenclmel, A which are mysterious when viewed autonomously in linguistic terms (e.g. explicit performatives, self-referring expreb sions, indirect speech acts, assertive or interrogative...). I take the present OP portunity to point out that although there has recently been a welcome rcnewal of interest among psycholinguists for speech act phenomena, they tend to underestimate the social complexity of these phenomena. Experiments on ‘Mental
88
6illesFauconnier
sample student groups interpreted by other academics are, in this case, invariably biased culturally, running the risk of interpreting social contingencies as psychological universals. A third area of my work may be mentioned in connection with the epistemological problems raised by the first two: syntact$ semantic, and pragmatic polarity. Wereagain ttere is ample evidence to shf Nthat abstract objects (implicational scales) are set up in discourse and operated on, yielding what often looks like independent structural properties of sentences. Furthermore, well-known laws of logic such as De Morgpn’s, Contraposition, or quantifier extraction, turn out to be only special cases of more general implication reversal laws which apply inconspicuously but very ordinarily in everyday natural language. The general conclusion of all this is, I believe, pretty clear: many conceptually novel, exciting directions of investigation are opening up in linguistics; their psychological implications are by no means negligible; they may alter our ways of thinking about language in fundamental respects, without undermiting the possibility of a scientifically rigorous approach. References Fauconnier, G. (1975) PragmaticScales and Logical Strxture. Ling. Znq., t&353-375. Fauconrnier, G. (1975) Polarity and Scale Primzip!:.z%eeedings of the Eleventh AegionolMeeting of
thechic~oLinguisticSociety. University of Chicago, Chicago, IL. Fauconnier, G. (1977) Polarit Syntactique et smantique. Ling. dnvestig.,1,1-37. FauccmW, G. (1978) Is there a Logical Level of Linguistic Representation? Theoret. J&g.., 5.31 -49. Fauconnier, G. (1979) Implication Reversal in a Natural Language. In F. Guenthner and S. .I. Schmidt @is.), Formal &mantis and Ragmtics. Amsterdam, Reidel. Fawxmier, G. (1980) PragmaticEntailment and Questions. In J. R. Searle, F. Kkfer and hf. Bierwisch @is.), Spetih Act Theory andlbgmatics. Amsterdam, Reidel, pp. 57-69. F~auconnier,G. (1980) Social Ritwl and Relative Truth in Natural Language.In A. Cicoorel and K. Kncrr ieds.), Advarrcesin &be&z1 Theory and Methodology: ?‘owatisan Intepvtion ofMicro and hfacro Apprvaches. London, Routledge, Kegan, Paul. Fauconnier, G. (1980) Mental Spaces--A Discourse Processing Approach to Natural Language Logic. In R. Zuber (ea.), Syntax and Semantics New York, Academic Press.
Cognition, 10 (1981) 89-95 @ Elsevier Sequoia S.A., Lausanne -‘-Printed in The Netherlands
89
Researchon context effects in word recognition: Ten years back and forth IRA FlSCHLER+ lJniversity of Florida
From a Skinnerian perspective, 1 am in the business of briefly flashing words an+ word-like stimuli at sometimes indifferent students and mea!;uring the time they need to make unnatural decisions about the flashes. But this is not the real business of cognitive psychology, any more than smashing tiny particles together is the essence of quantum mechsnics Such activities have been for many of us in the profession a primary method of trying to understand how people acquire, represent and utilize information. My particular flashing-and-measuring activities have been concerned with the momentary, complex processes that occur during the few handred milliseconds that it takes Yor a graphemic stimulus to elicit a meaningful verbal response. I have been using an experiment11 paradigm commonly called “priming” to study the role of attention in word recognition and reading (Fischler, 1977a; Fischler and Bloom, 1979), the nature and speed of associations between words (Fischler, 1977b; Fischler and Goodman, 1975), and the relation of semantic and episodic word recognition tasks (Fischler, Bryant and Querns, Reference note 1). The central axiom of a priming aiperiment is that the ‘content’ of a psychological event can influence response to an immediately subsequent event. The term was apparently borrowed from earlier demonstrations that particular free-association responses could be ‘primed’, or made more likely, by prior presentation of words semantically related to +he target response (Segal, 1967). The first demonstration of associative prin.inil; under ‘real-time’ conditions was reported by Meyer and Schvaneveldt ( 197 1), who founld that lexical decision latency (the time needed to decide if a string of letters forms a word) for pairs such as DOCTOR-NURSE was less than that for pairs of unrelated words. What made this marriage of priming and lexical decision particularly attractive, to myself and others was that the priming occurred rapidly-the latency of response to related pairs being less than a second, on avemge-and
‘*This paper ~8%pxspared while the author was on sabbatical as a postdoctoral Fellow in Psychology at Gallaudet College, Washington, D. C. Reprint requests should be sent to Dr. Ira Fischler, Department of Psychology, University of Florida, Gainesville, Flotida 32611, U. S. A.
40
I. Fischkr
seemed to dissipate almost as rapidly (Meyer, et al., Reference note :!). Associative priming was also shown to be a very replicable phenomenon, despite its modest magnitlJdc (typically 30-63 msec). This combination of robustness and temporal resolutir;n seemed ideally suited to an informationprocessing analysis of word recognition. Finding that the priming effect could be accounted for by a spreading-act&&ion process (Meyer, et al., Reference note 2) was in part responsible for the popularity of activation models developed during the past decade (e.g., Collins and Loftus, 1975), which led in turn to further use of the paradigm. Perhaps the mos: basic question we can ask about word recognition, from our information-processing perspective, is whether it is better characterized as a search through a ‘dictionary’of stored lexical entries (e.g., see Forster and Chambers, 1973), or as a detection of a signal by some sort of direct-access association of a visually presented word and its stored representation (e.g., Morton, 1969). Most theoretical treatments of priming have invoked a variant of Morton’s logogen model, which is a detection-threshold model of word retrieval. I know of no attempt to directly compare the two approaches experimentally, but since the choice could be seen as one instance of the more general ‘serial versus parallel processing’issue, I suspect that a final resolution is not likely (but see Wickelgren, 1981, pp. 22-24). Several recent models of retrieval have used both a passive, parallel detection process, and a more active search process, usually in that order (e.g., Anderson, 1976; Cecker, 1980; Juola, Fischler, Wood and Atkinson, 197 1). 0ne possibility that is seldom considered here-or for that matter, whenever we describe dichotomous alternatives in our theories -is that we could no more decide the ‘true nature’ of lexical retrieval than we could point to the ‘true nature’ of light as wave or particle. We may find that even so gross a distinction as search uersus detection camlot be resolved experimentally, and that in a fundamental way both views are correct.
A second important question about the nature ofword recognition concerns the role of ‘higher-order’ codes in various stages of word recognition. There is no doubt that the information provided by a sentence context, for example, can affect the speed and’accuracy of certain responses to subsequent words (Tulving and Gold, 1963; Schuberth and Eimas, 1977 ; Fischler and Bloom, 1979). The question rather concerns whether this effect is on the encoding and retrieval of the word itself, or on some later stage of comparison and decision. As with the search/detection question, this issue is a specific instance of a very general problem that appears in a number of areas: are sub$rocesses ‘autonomous’ (Forster, 1979) and therefore sensitive only to input from ‘below’, or can the flow of information be ‘topdown’ as well? Several recent and influti~~iiil models of cognition (e.g., Rumelhart, I 977) have explicitly incorponted krpdown processes in perception and comprehension But Forster
tintext effects in word recognition
31
(1979) has argn:ed that much of the work showing evidence in favor of topdown processing has failed to adequately distinguish the analysis of a lowerlevel code from decisions based on the output of that analysis. Forster himself presents results of a se& of converging studies which are reasonably consistent with an autonomous-firocessing approach. At one time, it appeared we had solid evidence that semantic priming affected the encoding of the targei word. Meyer, Schvaneveldt and Ruddy (1975) showed that changes in the visual quality of the target display interacted with relatedness. Application of Sternberg’s ( 1969) additive-factors logic implied that the presence of a related prime speeded visual encoding of the target. This of course would be strong evidence for the existence of ‘top-down’ processing. But the firmness of this conclusion has been tempered, primarily by McClelland’s elegant description of an alternative to Sternberg’s discrete-stage approach, McClelland (1979). showed that if output from earlier stages were allowed to ‘cascade’ into later stages before the earlicz: stage was complete, then the interaction of two variables in a task did not necessarily imply that a common stage or process was being affected. It is tempting to argue that the sometimes astonishing speed with which priming occurs (see Fischler and Goodman, 197Y; Fischler and Bloom, 1980) means that an early stage of target word processing is being affected, and so the priming effect must be top-down. But Forstcr:‘s argument applies here as well: although the elicitation of meaning from the priming stimulus may be very rapid, the influence of this information on the target cou1.d still be late in the response interval. (See Fischler and Goodman, 1978 for a discussion of this point.) Latency of priming effects, them, do not in themselves provide any evidence about sequence of stages or direction of information flow. Still, extremely rapid priming is interesting for other reasons. Since priming b.as been achieved with prime durations so brief that the identity af the prime itself is not available to the.subject (Carr, McCauley, Sperber and Parmlee, Reference note 3; Fischler and Goodman, 1978, Exp. 2; Allport, 1977), we are provided with a contemporary version of what a generation ago was called subliminal perception. ‘The phenomenon seems much less unsettling now thnn it did then (see Neisser : t967), and we are impressed but not really surprised by examples of ‘preconscious’ effects of stimuls meaning, wlhich after all are paradoxical only if perozption is seen as a unitary event Understanding tic .,J vt..ry rapid events has been made easier by the development of models distinguishing ‘automatic’ and ‘attentional’ processes. In one influent:iaI version, Pocmer and Snyder (1975) argued that certain overlearned, habitual processes could occur without effort. Using ,their framework, Neely (1977) was able to show that the priming effect had both attentional and automatic components. Much of the success of this and similar work was
due to the use of a cost-benefit analysis developed by Posner and Snyder, in which a neutml condition with no prime is used as a baseline to assess facilitatory and inhibitory components of the priming effect. Despite the successes of this technique for separating automatic and attentional processes, problems have arisen. Methodologically, it is difficult to design a truly “neutral” baseline condition that is equated with the prime conditions of factors such as alerting and processing-load characteristics (see Fischler and Bloom, 1980). The result has sometimes been that &hepattern of costs and b?neflts are uninterpretable (see, for example, Carr, et al., Reference note 3). A greater problem for the theory of automatic processing is that even when the neutral condition seems to be matched to the prime conditions on these factors, inhibition that apparently develops far more rapidly than Neely’s results suggest has been found with some regularity, both with single-word primes (Antos, 1979), and with sentence primes (Myers and Larch, 1980; Fischler and Bloom, 1980). The problem is that the presence of inhibition is supposed to imply attentional involvement. In the context of the Posner and Snyder model, this means either that inhibition can arise ‘automatically’, or that attentional processes ctin occur as rapidly as more automatic ones. In either case, some of the power of the original dichotomy is lost. Conversely, Spelke, Hirst and Neisser (1976) have shovvn that even quite complex tasks such as dictation can apparently become automatic, in the sense of not interfering with other simultaneous tasks. I agree with Myers and Larch (1980), who believe that the concept of allocation of resources in a more continous way (Norman and Bobrow 1975), and the methods associated with estimating the tradeoff of allocation between tasks (e.g., Navon and Gopher, 1979) will become increasingly important in the near future for understanding the relation between automatic and attentional processing. A more precise description of what we mean by those terms will also be an important development. Even in the absence of a wholly sailsfactory definition of automaticity then, the idea of automatic processing has been of great importance in understanding both priming effebts in particular, and the development of skilled performance in general (cf., LaBerge, 1976; Schneider and Shiffrin, 1977). One direction I b.ope to see this work take is a more thorough study of the relationship between early stages of leaning a task, where performwce is often mediated by conscious, attentional factors, and later stages where performance is more automatic. It could be that practice simply makes subcomponenta of a task and their interrelation more efficient, thus demanding less resources. But automatic performance may also be qualitatively different from more deliberate stages of acquisition. Specifying the nature of such qualitative shifts could have a major impact on the assessment of various teaching methods
Context effects in word recognition
93
fbr skills SU& as wdrng and second laitguage acquisition, where it is often assumed that methods producing initially superior performance will also provide ultimately higher levels of skill. One example of how work on attentional and automatic aspects of priming can be applied to skill development has been provided by Stanovich (1980). It appears that less skilled readers are, if anything, more dependent on contextual cues for word recognition than are older, more skilled readers. Stanovitch argues that younger readers rely more on context because their skill at automatically recognizing words in isolation is not yet well developed (cf. Perfetti, Golman and Hogaboam, 1980; Juel, 1980). The fact that skilled readers do not seem to actively anticipate the occurence of particular words in context (see also Fischler and Bloom, 1979; Mitchell and Green, 1978) suggests that an emphasis on this ability in reading instruction is misplaced. In the next decade, I hope to see increased use of some of the more powerful analytic methods recently developed, including speed-accuracy dynamics (Wickelgren, 1977), comparison of constructs across tasks (e.g., Hunt, Lunneborg and Lewis, 1975) and resource-allocation functions (e.g., Navon and Gopher, 1979). I also anticipate increasing use of psycho;~~~;:Jsiological measilrel;, in particular event-related cortical potentials, or ER:‘;, in the analysis of cognitive processes. Evoked-potential techniques offer an ui;usual combination of sensitivity, unobtrusiveness, and real-time analysis that has much to contribute to our field (see for example Donchin, 1980; Kutas and Hillyard, 1980). Our theories of memory and performance should also increase in generality and sophistication. I am less optimistic that we will have anything more to say about the nature of consciousness ten years aence. Cognitive psychologists walk a tightrope in using constructs such as central processing capacity and executive control without really grappling with the problem of consciousness and the homunculus except by analogy to computers. The qualitative, experiential nature of consciousness has also played no role in information processing theory, despite its importance in some recent problems in the philosophy of mind (see Fodor, 1981), Perhaps the integration of cognitive psychology with fields such as philosophy and artificial intelligence in the emerging discipline of Cognitive Science will provide a synthesis of experiential and information-processing aspecis of consciousness- - or at feast determine if this is feasible or desirable as a theoretical goal. Progress in experimental psychology is notoriously ill-defined, and my colleagues are quite divided on our accomplisnments (compare Tulviag, 1979), and Wickelgren, 1981). I suspect that most of us have felt pessimistic and optimistic by turns. 1 hope I have been explicit about areas of progress resulting from our collective work on word recognition, association and attention.
At worst, what we do is an intellectual treat, an entertainment for which someone is willing to pay us. But at best, our work offers the soundest path available to psychology for understanding the nature and capabilities of the Inind. Keffzrences Allport, D. A. (1977) On knowing the meaning of words we are unable to report: The effects of visual masking. In S. Dornic (ed.), Arrention und Performunce Vf. Hillsdale, NJ., Erlbaum. Anderson, J. (1976) Lunguqe, memory and thought. Hlllsdale, NJ., Erlbaum. Antos, S. J. (1979) Processing facilitation in a lexical decision task. J. exper. Psychol., 5, 527-545. Becker, C. A. (1980) Semantic context effects in word recognition and reading: An analysis of semantic strategies. Mem. Cog., 8,493-512. Collins, A. M.,and Loftus, E. F. (1975) A SFieading-activation theory of semantic Processing. f%ycho~. Rev., 82, 407-428. Donchin, E. (1980) Event-related potentials: A tool in the study of human information processing. In H. Begleiter (ed.), Evoked potenriulsin psychiorry. New York, Plenum. Fischler, I. (1977u) Associative facilitation without expectancy in a lexical decision task. J. exp?r. Psychol.: I;um. Percep. Perf. 3 18-26. Fischler, I. (1977b) Semantic facilitation without tassociation in a lexical decision task. Mem Cog., 5, 335 -359. Fischler, I., and Bloom, P. A. (1979) Automatic and attentional processes in the effects of sentence contexts on word recognition. .I. verb. Leurn. verb. Behuv., 18, l-20. Fischler, I., and Bloom, P. A. (1980) Rapid processing of the meaning of sentences. M?m. Cog., 8, 216-225. Fischler, I., and Goodman, G. C a(1978) Latency of associative activation in me:moty. J.exper. Psychol: Hum. Percep, Perf, J, ad~5-470. Fodor, 3. A. (1981) The mindqbody problem. Sci. Amer., 244, 114-123. F
Cmmxts effects in word recognition
95
R. W., and Ruddy, M. G. (1975) Loci of contextual effects in visual word recognition. In P. Rabait (ed.), Arten:ion and Performance V. New York, Academic Press. Mitchell, D. C., and Green, D. W. (1978) The effects of context and content on immediate processing in reading. Q. & eq&?r. Psycho?., 30, 609-536. Morton, J. (1969) interaction ot reformation in word recognition. Psychol. RPU., 76, 165-178. Myers, J. L., and Larch, R. F. (1980) interference and facilitation effects of primes upon verification processes. &fern Cog., 8, 465 -4 14. Navon, D., and Gopher, D. (1979) Qn the economy of the human information-processing system. Psy. chol. Rev., ff6, 214 -255. Ncely, J. H. (1977; Semantic priming and retrieval from lexical memory: The roles of inhibitionless spreading activation and Limited-capacity attention. J. exper. &vcAol: Gen., 106, 226-254. Neisser, U. (1967) Cognitive Psychdogv. New York, Prentice Hali. Norman, D. A., and Robrow, D. C. (1975) On data-limited and resource-limited processes. Chg. PryMWW, D. E., khwneveldt,
chol., 7,44-64.
Perfetti, C. R., Goldman, S. R., and Hogaboam,T. W. (1980) Reading skill and the identification of wordr in discourse context. Mem. Cog., ;,273-282. Posner, M. J.,and Snyder,C. R. R. (1975) Attention and cognitive control. In R. L. Solso (ed.), Informution processingund cognition. Hillsdate, Nl., Erlbaum. Rumelhart, D. E. (1977) Toward an interactive modelofreading. ln S. Dornic (ed.),Attenfion and Performances VI. Hillsdale. NJ., Erlbaum. Schneider, W., and Shiffrin, R. M. (1977) Controlled and automatic human information processing: I. Detection, search and attention. Psychol. Rev., 84, l-66. Schuberth, R. E., and Eimas, P. D. (1977) Effects of context on the classification of words and nonwords. J. exper. Psycho!: Hum. Percep. Perfi, 3.27-36. Segal, S. J. (1967) Prinlirig ofassccbtion test responses by differential verbal contexts. J. exper. Psychol. 74, 370-377. Spelkc, E., Hirst, W., and Neisser, U. (1976) Skills of divided attention. Cog., 4. 215-230. Stanovich, K. E. (1980) Toward an intcractivc-rompcnsatory model of individual differences in the development of reading fluency. Recd. Res. Q., 16, 32-71. Stemberg. S. (1969) The discovcqr of processing stages: Extensions of Dander’s method. In W. G. Ko$ter (ed.), Attention and Performance II. Amsterdam, North Holland. Tulving, E. (1979) Memory research: What kind of progress? In L+Nilsson (ed.), Peapectiveson Memory Reseurch. Hillsdale, NJ., Erlbaum. Tulviq, E., hnd Gold, C. (1963) Stimulus information and contextual information as determinants of tachistoscopie recognition of words. .I. exper. Psyclrol., 66, 319-327. Wickclgren, W. A. (1977) Speed-accuracy tradeoff and information processing dynamics. 4cfu Psycholo&a, 41, 67-85. Wickolgrcn, W. A. (1981) Human learning and memory. In hit. R. Rosenzwcig and L. W. Porter (eds.), Ann& Review ofPsyM~logy. (Vol. 32.) Palo Alto, Annual Reviews. Reference notes 1. Fischler, I., Bryant, K., and Ouerns, E. Associative priming in sesnantic and episodic memory. Presented paper, Southeastern Psychological Association, New Orleans, March 1979. 2. Meyer, D. E., Schvaneveldt, R. W., and Ruddy, M. G. Activation of lexical memory. Presented paper, the Psychonomic Society, St. Louis, November, 1972. 3. Carr, T. H., McCauley, C., Sperber, R. D., and Parmlee, C. M. Words, pictures, and priming: On semantic activation, conscious identification, and the automaticity of information processing. Manuscript in review, 1981.
Cogf&ion, 10 (1981) 97-101 @Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
97
Objects of psycholinguisticenquiry M. F. GARRETT” Massachusetts Institute of Technology
Analytic and experimental enquiry in psycholinguistics may be best understood in terms of a sin@e general qilestion, simply put as follows: Is human language‘merely’ the exercise of our general cognitive capacities in aid of communication -or, are there principles of mental representation and process which are specific to natural language?
Most of contemporary linguistics answers ‘no’ and ‘~6s’to the two halves of this question; a fair bit of psychology and artificial intelligence reverses the order of those responses. Though it might be instructive to ask winy, I will content myself with noting two aspects of the different approaches to language study fostered by this, difference in viewpoint. These are, first, a different expectation abobi the role of nonlinguistic background knowledge (‘context’) in sentence processing, and, second, a different expectation about the effects of task and modality, on langG?gPprocessing. These two points are related. A basic, oft noted, and little understood fact of real-time language processing is its context dependence. The issue is what we shall make of that fact. Incontrovertably, the interpretation of sentences is affected by situational variations in the availability of a grab-tag of types of information which are not part of the, normal inventory of grammatical structures. The problem that one faces is the simultaneous adjudication of claims for such influences with accounts of the phonological, syntactic and logical form of sentences- accounts which make no mention of the attitudes, beliefs, emotional states, age, sex, marital status, or educational background of speakers or listeners (all of which are in somebody’s grab-bag). Those who are Qersuaded of tP existence of a language faculty are so because of the existence of a theory of the distribufional and logical properties of Eentences which abstracts from the general knowledge of language users. Hence, they prefer language processing theories which provide a pristine account of grammatical and logical structure. If normal language processes do not reconstruct the representational types of grammatical theory, it removes one of
*Reprint requests should be sent to Merill Garrett, Psychology Department, MIT,Cambridge,Mass. 02139, U.S.A.
98
M. F, Garreit
the principal justifications for such theories-namely, that they represent the speaker’s knowledge of language structure. Those who doubt the existence of a language faculty tend to question the success of theories of sentence form. Hence, they are relatively indifferent to the consequence of introducing pragmatic variables into the computational models of sentence comprehension and production in ways that compromise the identification of the target representations of stizh models with specifically linguistic representations. Contextual constrtint is introduced as available, wherever potentially useful. By coatra,$, processing theories which identify a language faculty introduce contextu:ll cc ,lstraint as a means of selection among *competing representations provided by independent processors whose domains are the structural types of linguistic theory (note that there is no neces?;ary set of temporal claims here-contextual ‘expectations’ may be established prior to the elaboration of the structures among which they will select). The latter approach is modular in a way the former is not, for it restricts the introduction of contextua! constraint to the endpoints of the various computations based on linguistic types. These contrasts also suggest a different expectation about the effects of modality on language processing. Insofar as modality differences may affect the characteristic time course for the display of given sorts of information\, a contest centered view of language processing should expect corresponding processing changes. If the normal availability of a given information set determines which basic processing steps are taken, there should be modality specific properties of real-time language processing which reflect variations in the reliability and temporal availability of the informatiord which determines sentence interpretation. Alternatively, one may suppose that a basic determinant of real-time language processes is the relation among the information types in the domain of structures to be fixed. Hence, those relations will dictate the way in which information is sought in the physical Tignal and brought to bear on the analysis. A modular approach places limits &.\ponthe use of a given fact as a function of its type because interaction patterns are goverGed by type rather than by situational utility. Of course, these two principles are not incompatible in principle, and dobibtless both are correct. However, the extent to which one attends to each will be affected by one’s conviction that there is indeed an informational structure based upon sent.ence form wl- L.his distinct from the conceptual structures of our general knowledge of the world. Given the views I have sketched, what are the most revealing contexts for their evaluation? I believe it will be in a comparison of the organization of computational systems in different processing domains, specifically, in comparisons of language production and language comprehension systems,
Objects of psycholinguistic enquiry
99
on the one hand, and in comparisons of auditory/vocal language systems with visually based systems on the other, To illustrate briefly: Production versus comprehension
We know these systems are intimately connected in diverse ways-we monitor what we say for sense and form and we do so in the light of not only our own communicative intent, but also in the light of our judgment of listeners’ knowledge and capacity; mareover, the structural targets of the two systems must have in common at least a representation which unambiguously specifies the meaning of the utterance exchanged. Nevertheless, one might argue that the two systems have different ancillary rel;ponslbilities arising from differences in their primary tasks of interpretation and pronunciation. One might be tempted to say that compreheilsion systems can play fast and loose with sentence fom so long as interpretation survives -an assumption of contextual detemination by non-linguistic information types IS an essential feature of such a view, for it purchases the power to bypass or truncate computations of sentence form based on phonetic and syntactic features of utterance through substitution Qf inferentially dictated analyses. By apparent contrast, the production system must get nuances of form right and in detail, for otherwise one will sound peculiar-’ i.e., not as a native speaker, On these grounds, one might expect, if the thesis that contextually based inferential structure may be dyntzmically substituted for form based structures is correct, to find a closer relation
between theories of sentence form (i.e., grammars) and theories of real-time language production than that between theories of form and theories of real-time language comprehension. My own expectation is that the contextual thesis as stated will prove incorrect and that the relation between theories of sentence form and processing theories will be substantially the same for both comprehension and production. Audito.ry versus visual s: $ems: I. Readingfwriting versus listening/speaking
Here again, one fiLs ~,.~ny reasons to expect similarity in the two systems--reading is learned after and via the auditory/vocal system, and alphabetic writing systems are encodings of the sound structure of spoken languages. But- it is bruited about that the auditory signal has prob!ems the wit, segmentation and ‘data quality’; the physical visual one does not -to evidence for word boundaries is deemed, on such a view, to be less well developed in the auditory signal, and in general, th.e acoustic data are taken to be, on average. more degraded as representative OPphonetic types than are the visual data of orthographic types. Hence, one might argue, more context dependence may be required for processing of spoken language than for
processing of written language. To which one might reply: the auditory signal has a data base only .dimly reflected in orthography-namely prosody, and ii; theoretical structure and role in procl, ssing is even less well understood than the poorly understood roles of lexical structure and meaning relations. It should be clear that despite the intimate relation between sound systems and alphabetic systems, there are plenty of reasons to look for significant divergences in the information processing structures for languages in the two modalities. If one finds them, well and good-we will be on our ‘way toward finding out how the quality and temporal avar. at ‘-2~ of information affects the organization of processing systems. If one does not fird major differences in the two types of systems, the moral is fairly clear: The informational structure inherent in the language is a more signiticant determinant of processing organization than is variation in the physical manifestation of that structure. Auditory versus visualsystems: II. Visual/manualsign wsus auditory/vocal 1 speech One might object to the moral drawn above on the grounds that the strus ture of a language may indeed shape its processing systems, but that this is true because the influences of modality have been built into the structure over the evolutionary history of the language-hence, written versions of spoken languages are rrirt good test vehicles even though there may be demonstrable contrasts in the informational displays peculiar to the two modalities which ought, on a priori grounds, be expected t .I give rise to significant processing differences. Fortunately, this issue may be addressed more powerfully by a Comparison of signed languages of the deaf with spoken languages of the hearing. Recent study of, e.g., American Sign Language (ASL) shows even the hardened skeptic that it has the expressive power of a spoken language. Current research on its structure provides many indications that its grammar may be profitably compared with that elf spoken languages, for there are detailed, form dependent contingencies which seem analogous to the phonological, morphological and syntactic structures of spoken language. Clearly, if the ultimate development of grammars for s,gn shows them to be so similar to that of spoken language that they may be assimilated to the same general theoF/ of grammar, it will seriously compromise the objection raised to interpretation of a possible convergence of the processing systems for reading and writing, for that objection requires that t..lere be a significant impact of modality upon language structure. More powerful still would be the implication af correspondences or differences in the processing systems for signed and spoken languages, given that we may make a detailed case for
Objects of psycholinguistic enquiry
101
or against the similarity of their respective grammars. For example, if the grammars are different, we may test the hypothesis that the organization of processing systems is determined by the organization of grammar by looking for differences in processing for sign an;i spoken language ,Aich correspond to differences in their grammars. If the grammars are similar, the complementary prediction holds. If we were to find that the processing systems are different and that the grammars of sign and speech are similar, we would be forced to amend the view that grammatical structure is a primary determinant of processing structure. The thesis of modularity is, of course, tied in with these predictions, for if knowledge types may interact in unconstrained fashicn, there would be no reason to expect that two grammars which draw the boundaries of their rule systems differently, should therefore yield different processing systems. The sorts of observl;tions noted above are certainly not the only relevant ones. Such observations would, however, allow us to evaluate the impact of mod.ality specific features of information display on the development and exercise of processing systems both of which are in the service of a common goal-that of face to face communication. With detailed theories of the linguistic structures underlying performance in each domain, and with detailed information about the time-course of the processes which recover such structure, we might evaluate the effects of grammatical organization on processing organization, and the effects of modality of exercise on each of these.
Cognition, 10 (1981) 103-114 @Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
103
IMaturaMmal determinants of languagegrowth LILA R. GLEITMAN” University of Pennsylvania
Investigation into language learning durkg the last decade reveals striking uniformities in the acquisition process: under widely varying environmental circumstances, learning different lang lages under different conditions of motivation and child-rearing practices, and with different individual talents, all learners in the normal range acquire their native tongue in about six years. Moreover, there are into-r:?sting similarities in the course this learning takes. A few isolated words appear. + about age one, the learner speaks two and threeword sentences at about ag two, simple but essentially grammatical sentences at about three, and by four years or so sounds much like an adult speaker. In the succeeding few years, some complex devices of derivational morphology make their appearance and some item-specific information (e.g., learning that the past tense of hit is ‘hit’, not ‘hitted’) gradually falls into place. The same pattern of learning has by now been observed in many linguistic communities. How are we to account for these uniformities in learning pattern? The dominant position in deveiopmentaa psycholinguistics puts most of thL burden of explanation on the environment. It is generally agreed, regardless of theoretical persuasion, that this environment consists of hearing utterances of the language in the presence of relevant extralinguistic information. The child% task is to construct a system, based on these samples, that projects to an infinite set ,,f sound/meaning pairs (for this analysis, see Chomsky, 1965). But the problem of explaining learning, given these framework prir ciples,. has usually seemed more awesome to those who traffic daily in lin
guistic phenomena than it does to developmental psycholinguists. One difficulty is understanc%g just how the extralinguistic context aids the learner, since a given scene or event in the world can be described by many different sentences. For instance, scenes suitable to The cut is on the mat are also suitable to The mat is tender the cut and Get that cat offmy new mat. *I thank Scott Weinstein and Elizabeth !Spelkefor many helpful comments on this paper.Current support for the reported studies comes from grant #12-Z from the Nat3onal Foundation for the March of Dimes and from ##PO1HD 10965+4 from the Department of Health Education and Welfare. Reprint requests should be sent to Lila Clcitman, Department of Psychology, University of Penrwlvania, 3815 Walnut Street, Philadelphia, PA 19104, U.S.A.
104
LilaR. Gleitmun
Thus there is a long distance between meaningfully interpreting a scene and catching just how a heard sentence relates to ht. Another problem has to do with the variety and complexity of language forms to be learned. An account in terms of inductive generalization from samples seems possible in explaining how we learn The cat is on the mat from having heard The dog is on the carpet. But the account loses plausibility when we confront the perr’ormances within the reach of even the most ordinary individuals who learn a language. My current favorite example is the opening sentence (of two) in a recent letter to TV Guide: ‘How Ann Salisbury can claim that Pam Lauder’s anger at not receiving her fair share of acclaim for Mark and Mindy ‘s success derives from a fragjle ego escapes me.’ Among othler things, this writer knows that the first 2’7 words determine
absolutely that the 28th ends with an s. These are the kinds of performance tha.t a theory (;jf acquisition is ultimately responsible to. Descriptions that capture the:m (linguistic theories) seem so abstract that we are repelled by thl;: idea of assigning these to the learners. But what is the alternative? Usually the alternative taken in language acquisition research has been to account for The cat is on the mat and then cross one’s fingers. This is, anyhow, how I read much of the empirical literature on this topic. To be sure, very recently there have been serious attempts to model the acquisition process in its real glory (see, for example, Wexler and Culicover, 1980). Our own tack has bjeen to study this process in vitro, so to speak, so as to set realistic preconditions 0~1such models. During the last several years, we have looked at the environment requisite to 1ang:uagelealming. It is self evident that languages are learned partly ‘from the ou,tside in’ but I have been led by our investigations to suppose that much of the burden of explanation has to hexborne by innate capacities and dispositions in the learners, rules and representations specifically relevant to language (for discussion, see Chomsky, 1980). Therefore our aim has been to try to disentangle the internal and external resources the child is recruiting for the langua$:e learning task. Though our results are fragmentary and impoverished in various ways, we believe we have identified subcomponefits of language that are more or less environmentally influenced at particular learning periods. The studies have been cast in terms of the framework principle mentioned earlier: language is learned by axposure to utterancclinterpretation pairs. Certain populations of learners allow us to see what happens when one of these environmental factors is changed. Some learners receive less or inferior information about the utterances; other learners receive less or inferior informatidln about the contexts. More recent and planned studies are of
Matumtional determinants of languagegrowth
105
populations who are different in current mental state (learners whose ages vary) or final mental state (retardates). The question is how learning survives such changed environments and capacities. Varying the languagesamples There seems to be quite general agreement that a learner exposed to random samples of the sentences of a language would be unable to converge on the 1anglJage.The main difficulty is the richness of the incoming data, which would seem to support so bewildering a variety of generalizations (including wrong and irrelevant ones) that we would expect learners to vary extremely in the time at which they hit on the grammar.’ But this accords ill with the real facts, that all normals learn in approximately the same time period. The dominant response to this problem among developmental psycholinguists has been to suppose the environment provides detailed support for learning by ordering the input utterances. We have called this the ‘Motherese hypothesis’ (see Snow and Ferguson, 1977, for many papers adopting this position). It holds that caretakers present linguistic information in a set sequence, essentially smallest sentences to littlest ears. There is no doubt that adults speak differently to children than to adults so the utterances heard are not random selections from the adult language. The utterances to youngest speakers are very short, slow in rate, and the like. Perhaps this natural simplification from caretakers (whatever its motivatian) plays a causal role in learning. A number of theorists have pointed out that this notion looks less appeti zing when examined in detail (Chomsky, 1975; Wexler and Culicover, 1980). The narrower the range of available data, the more hypotheses available to account for it, and hence perhaps the harder to get to the right answer. Nonetheless, there is some surface plausibility to the idea that the mother first teaches the child some easy structures; when she observes from his speech that these; have been learned, she:moves on to the next lesson. To help this go though, we would have to grant the caretaker some implicit metric of syntactic/semantic complexity so that in principle she could choose judiciously the sentences good to say to learners. But here too there are some initial supportive findings: caretakers’ speech changes, to some degree, in correspondence with the learner’sage, In our studies of the Motherese hypothesis, we first collected extensive samples of maternal speech to 15 children aged one to 2% years. To our dis‘It is usually assumed, based on fmdhgs from Brown and Hanlon (19701, that learners are rarely aided by corrections of their tiwed or incomplete attempts.
106
Lila R. Gleitman
may, the properties of the maternal speech did not seem promising as aids to learning the forms of the language (see NR,:wport,1976). The mothers’ speech
forms were not canonical sentences very often and they were neither uniform syntactically, nor more explicit in how they mapped onto the meanings, than the sentences used among adults. Merely, they were short, and it is not clear how shortness alone will help the learner converge upon the syntax of the language. IJndaunted, we considered that some less obvious properties of the maternal speech might be aiding tPe learning. To study this, we revisited ,the original mother/child pairs six months later (Newport, Gleitman, and Gleitman, 1977). Analyzing the child’s speech at these two times, we were in a position to compute growth scores for each child, on many linguistic dimensions. The question was which properties of the mother’s speech at time one had predicted the child’s rate of growth, on each measure. One interesting outcome of these studies was that a number of dimensions of child learning rate were utterly indifferent to large differences among the mothers. For example, the child’s increasing tendency to express predicates (as verbs) and their obligatory arguments (as nouns) evidently is a fact predictable from maturation, not from the particular speech forms presented. On the contrary, in the age range studied, the childt’s progress with the closedclass morphology and functions (the ‘functors’, in ordinary psychological parlance, the little words and affixes such as CLLPI, have, 4) was a rather strict function of maternal speech style. For example, almost all the variance in rate of learning the English auxiliary items is predicted by the preponderance of yes/no questions in maternal speech. The effect of these questions is to place the auxiliary in first serial position, with stress (e.g., ‘Can you pass the salt?’ rather than ‘You can pass the salt’). Summarizing, certain univelsal properties of natural languages (expressing the predicates and arguments of propositions, for example) emerged in the child at maturationally fixed moments, and were indifferent to the naturally occurring variation among mothers. But elements and functions of the closed class, for children in this age range, seemed to be closely affected by specifiable facts about the input. Even here, however, the caretaker exerts her infhrence only as the information she provides is faltered through the child’s learning biases. For example the serial position, but not the frequency, of mothers’ auxiliary use affected the learning rate. Unfortunately, while these studies preclude certain strong forms of the L3%therese hypothesis, they leave almost everything unresolved. First, the iimited effects of envjronment on language learning that we found may kc? attributable to threshold effects of various sorts, to the attenuated sam2;‘3 (15 middle-class American mothers of children in a narrow age range), or to the measures or analyses used. These complaints are fair even though they
Maturationaldeterminants of language growth
107
lose some force given the positive findings for the closed class sub-corngonent. Nonetheless, there was clear Impetus for looking at cases in which t!x child’s environment was more radically altered. We therefore next studied a population grossly deprived of formal linguistic stimulation (Feldman, Goldin-Meadow, and Gleitman, 1978). These were six deaf children of hearing parents who had decided to educate their children ‘orally’, by having them taught to vocalize and lip read. Accordingly, in advance of the planned training period, these parents made no attempt to teach a gestural language. More iniportant, these parents did not Il-now a gestural language, so they were not in a position to present the easiest er-:amples first, the’harder ones later. It has been observed that children in these circumstances develop an informal system of communicative gestures, called ‘home sign’. It was the genesis of this system that we wished to study. Though many questions arise about how precisely we could analyze this exotic communication system, it is fair to say that the interpretive puzzles we fazed are not materially different from those confounding the study of young Evlglish speakers, by adult English-speaking psycholinguists. In each case, one has to try to interpret the child’s messages relying heavily on their rea.l-wornd context of use (cf. Bloom, 1970). In doing so, one encounters thi: same perils and pitfalls as the language learner himself. We settled for using the methods traditionally employed in studying normal language learning. And we achieved about the same results, for early stages. These linguistic isolates began to make single gestures at the same developrnentai moment hearing learners of English speak one word at a time. Two and three-sign sequences, encoding the same semantic/relational roles, appear at the same age as hearing learners speak in two and three word sentences. To the (rough) extent that the words in these primitive sentences are serially ordered by young hearing learners according to these semantic roles, similar serial ordering of the same categories described the self-generated gesture system. It seems then that ii ihe environment prov;Aes no sample sentences, the child has the internal wherewithal to invent forms himself, to render the same meanings. These results beccme more interesting when compared to the findings mentioned earlier, concerning the hearing learners. To the degree that the propositional forms and meanings appeared in indifference to variations in maternal input, these same properties appeared at the same time in the deaf learners, exposed to no formal language input at all. The closed class subcomponent, responsive at this stage to variations in maternal input, did not appear at all in the signing of the isolated youngsters. The fast suggestion here is that the closed class is laid down later, and in a different developmental pattern than other properties of the language system (for Bescrip-
108
Lila R, Gleitman
tive details, see also Brown, 1973).2 This is so despite the fact that the environment makes ali these kinds of data available from the beginning (mothers do not leave out the closed-class words, even to the tiniest learners). A second suggestion is that the one subcomponent of the language system is more environmer;tally dependent than the other, and may not appear at all in some exposure conditions. This is the interpretation we drew from the two studies. However, at least one discordant finding suggests this last claim may be too strong. .4 fascinating line of research (Sankoff and LaBerge, 1973) concerns the process of language formation: pidgins, and their creolization. This work suggests that the final (phonological) steps in creating a closed class morphology may be carried out by five to eight year old youngsters Iearning a pidgin as their first language. The pidgin to which these learners are exposed contains only impoverished closed class resources. The learners refine and expand these resources materially, changing the, pidgin into an elaborated language. This raises the possibility that the exposure-dependent findings we have might be limited to an early developmental period (as has been suggested by Furrow, Nelson, and Benedict, 1979, who have replicated some of these findings but interpreted them quite differently). The environmental effect might be only on rudiments-items aud substructures-af a system whose later growth is internally driven3 The clearer facts are that the closed class appeals late in language growth in a distinct developmental pattern (see Gleitman and ‘1anner, forthcoming, for a full discussion). This happens if it is modelled frcm the begirzring (as for learners of English) and sometimes even if not modelled (as for the learners of a pidgin). As for the isolated deaf learners, it is hard to say whether these elaborations were forever unava.ilable to them owing to their deprived circumstances, fox their later development is contaminated because overlaid with fragments of English and of formal sign language. We do kn0.w that at early s’rages, when non;lally circumstanced learners do not render most closed class items in speech, these isolated individuals give no evidence of devising them. Varying the interpretive inf omation
Currently, Barbara Landau and I are studying language learning in blind children. This study was conceived as the other side of the language learning zFindings from Bradley, Garrett, and Zurif, 1979; Marin, Saffron, and Schwartz, 1976; and Kean, 1977 suggest also that the clawed class is impaired differentbUy in various kinds of aphasia. Fromkin, 1971 and Garrett, 1975 sugges, also that it is recruited differently in speech per’brmance. ‘I am grateful to E. Spelke for some of the interpretations here.
Matumtional determinants of language growth
109
coin. In the studies I’ve mentioned, the child was in some ways deprived of information about language forms. What happens if the child is deprived of some opportunities to map heard utterances against their interpretations in the world? Surely, the blind learner suffers some such deprivations. Though he car hear, and touch objects, so can a sighted learner. A popular claim in the hterature holds that a chi! j learns which words refer to what (and hence their meanings) because, as he listens, he follows his mother’s gaze and pointing gestures. Even supposing (falsely) that the mother of a blind child names ,objects only when the child is holding them, in what sense could this be equivalent to gazing and pointing, in directmg reference-making’? Jn the light of these limitations on the blind learners’ opportunities to discern the referents of many heard words and sentences, we have been surprised to discover that blindness need not delay language learning. Two of the three subjects we studied were premature and one was full term; all were blind from birth. All three acquired language skills at about the same time as their appropriate premature and full-term sighted controls, on the measures traditionally used in language-learning research. Particularly curious is that the blind learners acquire about the same vocabulary items as sighted children, at the same times. Apparently, receiving different, and less, interpretable information has no dramatic effect on overall acquisition rate. Some details of the blind fzhild’s learning are quite interesting. We experted to find the largest differences between blind and sighted learners i*r acquiring the visual vocabulary,, words like iook and set, for here the information base is maximally different from the normal. However the blind child uses these words about as early as do sighted children. (In the education littirature, this fact has sometimes been taken as a sign of ‘loose thinking’ by blind children.) The meaning3 seem quite appropriate, though of course these terms map onto a different world. A sighted child told to look up will raise its face, but the blind child raises her hands. But it is not that the blind child mistakes look for touch. For one thing, she responds to ‘Touch the doll, but don’t 108~ at it’ by a tap or scratch on the doll, and then to ‘Now you can look at it’ by exploring it manually. On this and other evidence, we think the blind English-speaking child has developed a distinction as made in French, between toucher and triter, between manual contact, and apprehension by manual exploration. Our studies suggest that the caretakers use look to their blind offspring in a surprising way: much as parents do to their sighted offspring, i.e., to mean ‘look with the eyes’. To understand why they do this, try, with some acquaintance, to alter or restrict the circumstances in which you say look or see for about five minutes. I think this result suggests that despite the inismatch of maternal usage with objects and events the child can consistently interpret, the child interprets (better: comman-
deers) the term book to mark an important conceptual distinction: it renders percelption in the dominant mode through which this child apprehends the world,, while touch renders sensory contact with the world, in that mode. This is the level of description at which, it seems to me, look means much the same to this blind child as to a sighted child. The information for constructing this description is marvelously different for the two; however. Agai.n this suggests to me that an explanation of what is learned has tu rest materially on the reyresentations the child brings into the learning task. Varying the endowment of the learner
The studies I have sketched are consistent with a maturationally driven acquisition process, with progress relatively independent of exposure time or type. Early findings from ongoing studies that look at learners in different mental states lead me to consider a strict stage hypothesis for language learning. At the earliest stage, the learner is incapable of repi,esenting language. In a succeeding period, he represents the incoming language stimulation as a ‘conceptual’ system in which each sentence approx;mates a single proposition, the units are semantic/relational ones (‘agent of the action’, and the like) and the closed class structure is impoverished (see ,‘:obin, 1980). Finally, the normal learner jumps over to a different state, one of whose outcomes is a differently organized language, one that includes grammatical relations that are partly inldependent of semantic categories and relations. This is a sort of egg/tadpole/frog hypothesis, in which the mechanism of change is more like metamorphosis than learning. I am suggesting that reworking old data, not only taking in more new data, is a major component describing the learner at successive developmental points. Evidence will come from learners with varying mental endowments (c-rtain retardates) and those whose exposure to language comes at different times in maturation (premature children, second language learners). Ann Fowler, Rachel Gelman, and I have been looking a: I- guage learning in a small group of Down’s Syndrome adolescents. They were selected for homogeneity on several measures of cognitive funztion (e.g., MA about six years) ad an anchor measure of language function (mean length uf utterance, MLU 3-O-3.5). Our hypothesis was that they would be forming inductions over different, and shallower, units than normals. If so, they would have more to learn than the normals, accounting for the fact that they knew after 12 years or so only as much as their controls (normal 2% year olds with the same MLU’s) knew many years sooner. Notice that some hypothesis is necessary to descrille the retardates’ late learning, beyond saying they are ‘:dOW learners’.
Maturationalde,erminants of language grow.91 111
But as usual our hypothesis was tvrong. In excru-iating detail, the language of the DS individuals was identical to that of the normal c:~:ntroLs. Based on the earlier studies, we had expected selective difficulty with the closed class, but there is no evidence supporting this in our findings. However, we are no longer sure that it took the DS individuals inordinately long to learn what they did learn. An initial look at a population of seven-year old DS children who, on the same measures, seem to be younger versions of our adolescent population, suggests tlrat they already know just what the adolescents know about language. This finding is very tentative at present. But it raises the possibility that these children may be able to use linguistic data rather efficiently.4 Given the fact that language onset is late for them, they may have learned in the period from four to six years much what a normal learns in the period from one to three. Defensible evidence must come from longitudinal studies, which we now have underway. But ‘I am conjecturing that these individuals may not be best characterized as inefficient learners, slow to form generalizations. Rather, they arrive late at t’rte stage that allows learning to begin, and the learning stops early, owing to failure to reach the mental state in which normal learners rework the data, achieving the mature grammar. Some evidence consistent with this approach comes from a study Barbara Laudau and I are doing, of language onset in premature children. Results from a small initial sample suggest that onset time is predictable as time since conception (neu.rological development) not time since b&h (exposure time). This leads me, to ask whether normal learners first exposed to language data at later ages might not learn in qualitatively different ways. For instance, Lennebeig (1967) reported on young children who had learned some language but then seemed to lose all of it due to brain injury. The youngest victims of aphasia, as they relearned, seemed to pass through the usual stages of language learning, but more rapidly. The p attern of learning for somewhat older releamers did not resemble first-language learning. A likely source of further evidence, 1 think, is the character of second language learning in three to five year o~cl normals. Perhaps the early ‘conceptual language’ stage is not traversed during I :srr&lg by these older children for they are capable of the mature representations of linguistic data. This would help explain why a four year old foreign child, transported to America, seems to require only one year’s exposure to speak English like a native five year old (who has had five years of exposure). ‘Whatever the stable generalizations will be about the subsets of DS subje.:ts we are studying, these in no way are taken to reflect the retarded pOpUhtiOn as a whole, which vtiries in many ways. Even some DS victims achieve close to full language learning. Our interest is in characterizing ‘intelligence’ in carefully selected subpopulations.
112
Lila R. Gleitman
I hasten to repeat that these are predictions in terms of which we are organizing new work, not .fmdings. But if the predictions are supported they would speak to why it has proved less fruitful than hoped to look at child learners at two stages in dlevelopment, and then try to puzzle out how stage two was learned from stag,e one. The problem is a logical one. The language acquisition literature seems to tell us that two year olds have a ‘semantics based’ grammar. Thti supposition seems natural to psychologists who find it implausible to suppose that very young children are in possession of abstract categories such as ‘subject of the sentence’ that are semantically incoherent (an example is ‘how Ann Salisbury can claim that Pam Lauder’s anger at not receiving her fair share of acclaim for Mark and Mindy ‘s success derives from a fragile ego’). Rather, the child’s initial noun-phrases might be ‘agents of the action’. I am not as sure as some developmental psycholinguists that this semantic stage has been shown very convincingly (maybe it’s frogs all the way down). But assume for sake of argument that this will turn out to be the fact. If it is, it makes another plausible assumption untenable: that one moves from this star: to the mature stage in a way that preserves the categories and functions of the stage before-a continuous 1earnir.g sequence. The problem is that the semantic units are at best useless and at worst misleading for acquiring the grammatically functioning units of the adult language. For example, sentence subjects in the adult language, but also sentence objects, may play any of the semantic/relational roles. It is thus hard to see how the initial semantic scaffolding allows the later learning, except by dismantling it. It seems easiest to accomodate the findings about an early conceptual language b:r supposing it is superceded, not added to, at a later maturational stage. This could be owing to its failure to capture further data, but it could also be owing to qaalitative changes in the mental apparatus currently subserving acquisition. I have not meant in this discussion to minimize the very interesting class of problems that will have to be described by inductive processes, that secure the outcome that French children will learn French and Greek children will learn Greek. Serious and promising inquiries into these processes have come from Maratsos (1978) and Morgan and Newport (1981), among others. I have meant to suggest that inquiry into how learners represent linguistic stimulation to themselves is required to establish the categories over which induction takes place. It may be that different rules and representations become available successively, consequent on internal changes in the learners. In that case, for example, a model for learning the final (transformational, lexical/functional, etc.) grammar should not be responsible for predicting earlier language stages, from which it might be relatively independent. Whatever the correct theory, however, it is slear that language learning cannot be
Maturationaldeterminants of language growth
113
satisfactorily described simply as an environment-driven process. Rather, creative activities in the child, specific to language, do the lion’s share of explanation.
References BlOOWL (1970) Language development: Form and function in emerging grammars. Cambridge, Mass., MIT Press. Bradley, D. C., Garrett, M. F. and Zurif, E. G. f(1979) Syntactic deficits in Broca’s aphasia. In Caplan, D. (ed.), Biologicalstudiesof mental processes. Cambridge, Mass., MIT Press. Brown, R. (1973)A first language.Cambridge, Mass., Harvard University Press. Brown, R. and Hanlon, C. (1970) Derivational complexity and the order of acquisition in child speech. In Hayes, J. R. (ed.), Cognitionand the development of language. New York, Wiley. Chomsky, N. (1965) Aspects of the theDryof syntax. Cambridge, Mass., MIT Press. Chomsky, N. (197cj Reflections on kmguage. New York, Pantheon Books. Chomsky, N. (1980) Rules and representations. Behav. Brain Sci., 3, 1-6 1. Feldman, H., Goldin-Krsdow, S. and Gleitman, L. R. (1978) Beyond Herodotus: The creation of language by linguistically deprived deaf children. In Lock, A. (ed.), Action, gesture, and symbol: The emergence of language. Lo.tdon, Academic Press. Fromkin, V. A. (1971) Th non-anomalous nature of anomalous utterances. Long., 47, 27-52. Furrow, D., Nelson, K. and Benedict, II. (1979) Mothers’ speech to children and syntactic development: some simple relationships. J. child Lang., 6. 423-442. Garrett, M. F. (1975) The analysis of sentence production. In Bower, G. (ed.), Psychology oflearning and motivation(vol. 9). New York, Academic Press. Gleitman, L. R. and Wanner, E, (In press) Language learning: state of the state of the art. To a, spear in Wanner, E. and Cleitman, L. R. (eds.), Lunguage learning: state of the art. Cambridge. Cambridge University Press. Kean, M. L. (1977) The linguistic interpretation of aphasic syndromes: Agrammatism in Broca’s aphasia, an example. Cog., 5, 9-46. Lenncberg, E. H. (1967) 7% biologicalfoundations of language. New York, Wiley. Maratsos, M. (1978) New models in linguistics and language acquisition. I;: Halle, M., Bresnan, J. and Miller, G. A. (eds.), Linguistic theory and psychological rek’ity. Cambrragei Mass., MIT Press. Marin, O., Saffron, E, and Schwartz, M. (1976) Dissociations of language in aphasia: implications for li normal function. An. A! Y. Acad. Sci., 280, 868-84. Morgan, J. L. and Newport, E. (1981) The role of constituent structure in the induction of an artificial language. J. verb. Learn. verb. .rjehav.,20, 67-85. Newport, E. (1976) Motherese: The speech of mothers to young children. In CasteBan, N., Pisoni, D. and Potts, G. (eds.), CognitiveTheory, 2. Hillsdale, NJ, Lawrence Erlbaum Associates. Newport, E., Gleitman, H. and Gleitman, L. R. (1977) Mother, I’d rather do it myself: some effects and noneffects of maternal speech style. In Snow, C. E. and Ferguson, C. A. teds.), Talkingto children: language input and acquisition. Cambridge, England, Cambridge University Press. Sankoff, G. and LaBerge, S. (1973) On the acquisition of native speakers by a language. Kivung, 6, 32 -47.
114
Lih
R. Gleitmatt
Slobin, D. I. (1980) The repeated path between transparency and opacity in language. In Bellugi, U. aad Studdert-Kennedy, M. (eds.), Signed anll spoken language: biological constraintson linguisticform (Dahlem-Konferenzen). Weinheim, Verlag Chemie. Snow, C. E. and Ferguson, C. A. (eds.) (1977) Talkingto children: languageinput and acquisition. Cambridge, England, Cambridge University Press. Wexler, K. and Culicover, P. W. (1980) Formal priprciplesof languageacquisition. Cambridge, Mass., MIT Press.
Cbgnition, 10 (1981) 115-118 @ Elsevier Sequoia &A., Lausanne - Frinted in The Netherlands
11s
Cognitive systems, ‘folk psychology’, and knowledge 0. W. HAMLYN* Birbeck College, University of London
Present-day cognitive psychology suffers, to my mind, from the inheritance of behaviorism. It does so in the sense that the poverty of behaviorism has encouraged psychologists simply to supplement it by inquiring into the intervening cognitive systems that may be supposed to link stimulus and response, without sufficient consideration being given to whether those terms of reference are themselves satisfactory. It might be thought that such a claim is false on the grounds that the model on which much cognitive psychology relies is quite different, being derived from computing. Its concern is thus to isolate those functional systems which link input and output, such systems working on computing lines given the data-base supplied by sensory imput. It is no doubt historically true that this latter :nodel is quite foreign to behaviorist thinking-that, as Charles Taylor (1964) has indicated, having historically a much older source in 19th ecntury associationism. Nevertheless, it seems to me that there remains a tendency to think of the input and output in terms which are closer to those of‘ stimulus and response than they should be, especially given the fact that the terms “stimulus” and “response” are two of the most unclear and confused terms in psychology (Hamlyn, 1970). The suggestion that there is such a tendency may again seem an odd judgment in view of the plethora of theories of perception that exist-to confine attention for the moment simply to the input end. Yet, even in the case of sophisticated theories of perception, such as that of J. J. Gibson, there is a tendency to view what happens in perception as a function simply of the structuring of stimulation available tin the perceiver. That way of thinking also encourages the belief that all we need to consider from the point of view of cognitive psychology is the individual considered as a cognitive system in relation to his specific environment-something that has been true, historically, of the behaviorist apuroach also, as well as of approaches which have nothing to do with that, e:g., that of Piaget and his associates. For reasons that I shall come to directly, that too seems to me a mistake, The point which I wish to emphasize now, however, is that an oversimple view of input and output may encourage in turn an over-simple and
*Reprint requests should be sent to D. W. Hamlyn, Department of Philosophy, Birbeck College, UniMalet Street, Londorl WCIE7HX, England.
versityof London,
116
D. W.Hamlyn
premature identification of supposedly intervening systems characterized in terms of their functional role in relation to input and output so conceived. From the physiological point of view there would be nothing wrong with taking input as the stimulation of nerve-endings and output as muscular or other bodily movement, but to relate these via intervening systems in such a way as to provide the underpinning for anything that is psychologically interesting requires such a fantastic degree of physiological detail that the mind boggles. If what intervenes betwe.en input and output is to be such that it is characterizable in terms that are psychologically interesting the same must be true of the input and output thenrszives. They too must be characterizable in terms that are psychologically interesting. The one is not independent of the other. What, then, determines what is of psychological interest? I suggest that what determines what functional systems we are to look for as linking input and output characterized in psychologically interesting terms is what has been called “folk psychology” -the psychology of common sense that we employ every day of our lives. The psychological entity that is the human individual cannot be construed simply as a black box with enelagy input and output. If we thought that il could be so construed, then, given an input of energy in some form and corresponding output, we should have no idea of what sort of systems were to be found within the box such as to relate input and output. In fact we know the sort of thing to look for because of our ordinary, everyday understanding of the sort of thing a human-being is, psychologically considered. That cognitive psychology has to do with perception, imagination, attention, memory, thinking and so on is, as it were, a datum; it,is something that the psychologist brings to the situation from our everyday “folk” understanding of people. It is only in that way that we can have any general idea of what functionally determined systems are likely to be found within the “box” and how Lhe input and output are themselves to be conceived. Once given r3at, the experimental scientist can investigate what elaborations of the basic framework are necessary in order to deal with the empirical facts. But without that framework he is, so to speak, blind. If. however, we take folk psychology seriously in the way that I have insisted we must, we must also, if we are to undg=rstand cognition, take seriously some of those concepts which are essential to it. A central concept in this respect is that of “knowledge”. Memory, for example, is not just a matter of simple storage; it must at least involve storage of infomation, and that from the point of view of the individual concerned means in effect retained knowledee. It is in Grtue of that that the capacities that depend on memory function as they do. Thought, at least to the extent that it involves problem-solving,
CiigLtive sy Ttems, yolk pwhology ‘and knowledge
117
implies the application of knowledge, generally to get new knowledge; and it is plausibly to be maintained, as ii was as long ago as Aristotle, that all learning both involves knowledge and is dependent on prior knowledge. That a cognitive being is also a knowing being is thus cnrcial. That fact ought also to affect our conception of the input and Irlutput that are relevant for those interver.ing systems the function of which is to link input and output in a psycho”ogically interesting way. For if a memory system, for example, is really one which makes possible a certain kind of knowledge, then what supplies that system and what it in turn effects must also involve knob iedge in some way. Anu so it turns out to be. For the input-perception, as it surely must be conceived -does indeed have a connection with knowledge. For one of the roles of perception (though not the only one, since perception has non-epistemic, purely sensory and purely aesthetic features also) is to provide knowledge of the world around us. It is also mediated by knowledge. since perception is concept-dependent, and the possession of concepts is the possession of a form of knowledge or understanding of things. Analogous things apply at the output end, since what perception and the subsequent psycholr;gical functions mediate is not just any kind of bodily movement, but action; and action involves (not always perhaps, but typically) both skill and intention-notions both of which imply knowledge, knowledge of how to do whatever it is and knowledge also to the extent that what is done is done knowingly, as intention implies. Knowledge, however, presupposes the possession of a concept of truth or of something that plays the same logical role (Hamlyn, 1978). To be properly said to know something we must have got whatever it is right--whether the object of knowledge is a fact, a way of doing things, the sense of our own agency or whatever. But the notiond of truth, correctness or rightness are in a sense social notions. It is not that what is right, correct or true is socially determined; but rather that these notions imply that of a norm and that the appreciation of the force of a noLlfnis something that requires rjihers to impress it on us. It would make no sense to suppose that an individual could acquire an appreciation of -the force of a norm by himself. (A purely selfimposed rule is not really a rule, except perhaps in the context of rules of other kinds, as in effect Wittgenstein [ 19531 has shown.) All this entails that knowledge is not possible,. properly speaking, except in creatureswhich stand in something like a social or interpersonal relationship; and we ascribe knowledge to animals to t.he extent that they conform to or at least approximate to that condition. Hence, any adequate theorizing about cognition must presuppose that wha.tever is c(ognitive is social too. The social cannot lie brought in, as it is necessarily
118
f?. W.Handyn
with e.g., Piaget, as simply one aspect among others of the individual’s relation to the environment. A cognitive individual must already be social if it is to have any genuinely cognitive relationship with the environment, and the social cannot be simply added on to whatever cognitive relations with environment already exist. That does not entail that cognitive psychology, or developmental psychology for that matter, are simply branches of social psychology, since the relations to the social that are presupposed within the theoretical thinking of those different disciplines are themselves different. It does, however, mean that the framework within which cognition is to be viewed if it is to be viewed in a theoretically satisfactory way, must be enlarged from that of a system supplied with input and leading to
Cc,~lirion, 10 (1981) 119-126 @ Elsevier Sequoia S.A., L;tusanne- Printed in The Netherlands
119
Cognitive processesin readin ALICE F. HEALY” Yale University
My research is aimed at developing an understanding of the cognitive processes involved in reading. I have concentrated my investigations on two fundamental questions: What are the units used in reading and how are they processed? These questions have been addressed by psychologists as early as Huey (1908/1968), who appreciated ‘that the reader’s acquirement of ease and power in reading comes through increasing ability to read in larger units’ (Huey, 1908/ 1968, p. 116). What do we mean by ‘units’? In the physical world, units range from subatomic particles to galaxies. Although psychological units are harder to specify, one of their distinguished properties seems evident: Once individuals have abstracted a unit, they no longer concern themselves with its constituent parts. For example, having perceived a musical cord, listeners no longer attend to its component tones. Taking this property as the definition of a unit, we can then confront the task of determining which units are employed. Towards this end, I have deve2ped a simple detection task that indicates which. units are used in reading text by revealing which constituents are ignored by the reader. Subjects are asked to read a passage of text and circle every instance of a given target (e.g.; the letter t). More errors are made on very common words, such as the, than on rare words, such as thy (Healy, 1976), and, the pattern of errors is further determined both by the structure of the search passage (Drewnowski and Healy, 1977) and by the subject’s reading ability (Drewnowski, 1978, 1981). Adam Drewnowski and I (1977) have proposed a model to account for the pattern of results found in this detection task. The gist of this ‘unitization’ model is that subjects tend to miss letters on familiar words like the because they process such words automatically in units larger than the letter without completing processing at the letter
*The researclt reoorted here is currently being supported by NSF Grant BNS80-25020 to the University of Colorado. I was supported by a Senior Faculty Fellowship from Yale University during the preparation of this article. I am Indebted to W. K. Estes for insightful comments about this article. I&prim requests should be sent to: Dr. Alice F. Healy, Department of Psychology, University of Colorado, Muenzinger Building, Campus Box 345, Boulder, Colorado 80309, U.S.A.
level. II am currently in the process of extending, elaborating, and testing this model. By employing the detection task and related procedures, I hope to discover the size of the units used in reading words of various types, to investigate the determinants of reading unit size, and to determine the information processing rules applied to these reading units. Background Although most investigators agree that unpronounceable sequences of letters are processed letter by letter, investigators do not agree on which units are use3 to pro:ess normal prose text: some (e.g., Gough, 1972) have argued that *words,like unpronounceable letter sequences, are processed in terms of letter units, whereas others (e.g., &good and Hoosain, 1974; Smith, 197 1) have argued that words are read as single-unit pattern*. Still others have proposed reading units of intermediate size, such as spelling patterns (Gibson, 1965) and syllable-like voca& center groups (Spoehr and Smith, 1973). The lack of agreement concerning the size of the units used in reading prose is probably due to the fact that the units employed depend upon the subject’s reading skill (Gibson, 1971; Samuels, LaBerge, and Bremer, P978) as well as the nature of the reading materials and the task demands (Estes, 1975). Moreover, it seems reasonable that a particular individual, given a particular passage to read, would employ units of different sizes not only within the passage but also within a single phrase or even a single word. For example, the individual may read the stem of the word evolvingin terms of letter units, but may read the suffix as a single unit (Drewnowski and Healy, 1980). Because of the likely existence of multiple units, a number of theorists have proposed models of the reading process that include several types of units; see, for example, Lstes (1977), Gibson (197 l), IaBerge and Samuels (1974), Massaro (1975), Rumelhart (1977), and Smith (1971). My concep tion of the reading pro’cess borrows heavily from these modeis, although there are some important differences. Unitization model
The model proposed by Adam Drewnowski and myself r:1977) includes reading units at various levels of the linguistic hierarchy, suc.h as letters, syllables, words, and phrases. Our formulation, like Rumelhart”s (1977), postulates that subjects process text in parallel at the various levels available to them and that fa;niliarity with a unit at a given level facilitates its processing. 1.~:
Cognitive processes in reading tcx t
121
addition, we proposed that once a unit has been identified at a given level, the subjects proceed to the next unit at that level without necessarily completing the processing of units at the lower levels in the hierarchy. For example, consider the phrase the porcupine, which is unlikely to be processed as a single unit because of its infrequency in the languagt:. Once the subjects have identified the word the at the word level, they move on to the word porcupine, which they processat all levels in parallel, without necessarily completing the processing of all the letters (t, h, and e) within the word the. By making the additiolial assumption that identification of a given unit requires that it be completely processed,’ we can use the detection task, in which subjects identify targets at a given level, as a tool to determine which units have been completely processed in the course of reading text.
The detection task A numbe; of earlier investigators have used letter detection paradigms of various types to demonstrate that words are read in terms of units larger than letters (see, e.g., Johnson, 1975; Wheeler, 1970). However, there has been considerable dispute and confusion corcernirig these studies (see., e.g., Baron, 1978; Maaaro and Klitzke, 1977; Sloboda, 1976; Thompson and Massaro, 1973) and the analogous studies dealing with speech perception (see, e.g., Healy and Cutting, 1975; McNeil1 and Lindig, 1973). In fact, investigators have argued that the basic units of reading are larger than letters both on the basis of superior performance on letter detection in word contexts (e.g., Wheeler, 1970) and on the basis of inferior performance on letter detection in word contexts (e.g., Johnson, 1975). Various differences in the designs of these studies have permitted such seemingly contradictory results and have been largely responsible for the confusion. For example, there are critical differences across studies in the dependent variables considered (errors versus response lal,ncies), display time (see Johnson, 1975), and placement of target cue (precue versus.postcue; see, e.g., E&es, Bjork, and Skaar, 1974). In my research I take another route towards settling the question of whether reading units are ever larger than letters. I manipulate the variables that I expect to influence the size of the reading units-such as position in to determine whether letter detecthe text and frequency in the language’ Although it is reasonable to assume that identification of a unit requires the complete processing of it, the converse does not necessarily hold. Subjects may complete processing of a unit and it may influence their subsequent behavior, but the unit may not be accessible for other purposes such as overt identification (cf., Healy and Levitt, 1980).
122
A. F. Healy
tion is influenced by these variables. Different strategies may be employed in different detection paradigms, thereby leading to different results (see Jotnson, 1975, for a similar argument). I have chosen to employ the detection paradigm introduced by Corcoran (1966), because in that paradigm deteeLion occurs in a normal prose context.
The retjundancy hypothesis Although the unitization model can account for the large number of letterdetection errors made on the very common word the, other apparently reason;;ble e:iplanations of this effect have also been proposed (see Healy, 1975). The most popular alternative :xplanation is the ‘redundancy’ hypethesis (e.g., Corcoran, 1966). There are several different kinds of redundancy (Smith, 1971), but only semantic and syntactic redundancy are referred to by this hypothesis. Specifically, this hypothesis postulates that subjects miss the t in the function word the because the surrounding word context leads them to anticipate the occurrence of the word the so that they need not scan that word. I have provided several lines of evidence indicating that the redundancy hypothesis is inadequate. In one experiment (Healy, 1976), I demonstrated that subjects made a disproportionately large number of letter-detection errors on the word the in a scrambled word passage. The word the was not predictable from the surrounding word context in that passage, so that syntactic and semantic redundancy could not be operating. In a more recent proofreading study (Healy, 1980), subjects read passage i of text and searched for misspelled words. Whereas subjects searching for target letters in passages without misspellings made most of their detection errors on the word the, subjects searching for misspellings made relatively few proofreading errors on the word ttle when it was misspelled. These results indicate that subjects do not skip over or give inadequate attention to the word the in reading prose. The data instead implicate the involvement of umlization, since processing in terms of word-level units should be disturbed for misspelled words. More direct support for the unitization model came from letter detection experiments (HeaIy, 1976, 1980) in which I compared common nouns to rare nouns of the same length and with the letter r in the ;same locations. Subjects searching for fs made more errors on the common than on the rare nouns, which is what would be expected if subjects are more likely to read frequent than infrequent words in units larger than the letter. Another type of support for unitization came from experiments in which the passage read
Cognitive processes in reading text
123
by subjects was ‘lped with every other letter in capitals (Drewnowski and Hea&, 1977; Healy, 1980). This manipulation led to a large drop in detection errors, presumably because the formation of reading units larger than th,: letter was disturbed by this unusual form of typing. Although syntactic and semantic redundancy i:Q doubt :qlay some role in these studies, there is clearly evidence for the involvement of the unitization process.
Reading units Although we are not yet able to specify precisely the units that are used in reading different texts, the detection task has already revealed some guidelines concerning the size and the nature of reading units under certain circumstances. Using this procedure, Drewnowski and I (1977) disco rered that common function words, such as the and and, are often part of reading units that include more than one word when they occur embedded in familiar phrases. More recently (Drewnowski and Healy, 1980), we found that whether a given letter sequence, such as -ing, is read as a unit depends on its location within the word, its linguistic function in the text, its frequency in the language, and its spatial predictability within a word.
Processing rules In nddition to disclosing information about reading units, the yrocc i.u-esI have been using have revealed information about the rules readers use to process these units. For example, in a recent study (Healy, in press), I asked subjects to read passages of text and circle every instance of a misspelled word. The misspellings in these pqssages were generated by replacing a single letter in a word with another one, and I examined the relationship between the original letter and the one substituted for it. The pattern or’proofreading errors suggested a specific set of information processing rules applied by readers to their visual representation of the text. In particular, my results implicated a hierarchical feature test, according to which readers give first priority to resolving the general shape, or ti-,llvelope,of the letters and second priority to discriminating additional visual features. In addition, a sophisticated guessing decision rule was implicated for the misspellings that do not alter letter envelope: Readers tolerate letter features that are missing but are intolerant of additional features.
124
A. I? Healy
Alternative experimental tasks
Although the evidence for these information processing rules is quite compelling in my experiments, these conclusions are not completely consistent with the ‘patterns of confusions found in studies involving tachistoscopically presented letters (see, e.g., Townsend, 197 1). However, Garner and Haun (1978) have demonstrated that the patterns of visual confusions obtained in a study of tachistoscopic letter identification depend on the type of percep tual limitation leading to errors. When short exposure durations are responsible for errors, a different set of confusions arises than when errors are caused by the addition and deletion of line segments in the stimulus. Because different types of perceptual conditions lead to different conclusions, it seems wise to investigate the perceptual conditions closest to those observed in the typical reading situation. The proofreading task, in which subjects examine photocopies of typewritten prose passages, seems to approximate normal reading conditions better than the other laboratory paradigms that have been used to investigate alphabetic confusions. One potential problem with the detectior! and proofreading tasks is that they do not involve reading alone. Rather, an additional requirement-circling a target or a misspelling-is superimposed on the reading task, and this requirement may change the nature of the processing units and strategies Fubjects employ. In analogy with the Heisenberg (1930/l 949) unsertainty principle, the process of making our expemiental observations may perturb the very phenomena we hope to observe. However, it seems somewhat unlikely that subjects, who have many years of experience with a certain set of processing strategies, will choose to employ a radically different set of strategies when the reading situation is altered to a small extent. In fact, Stroop (1935) has demonstrated that it is difficult for subjects to disengage their typical reading responses even when the laboratory tasks demand that they do so. Furthermore, Drewnowski and I (1977) have found that when the normal left-to-right reading pattern is impaired by presenting material in a list for.mat, rather than in the standard paragraph format, the number of errors in the letter detection task on the word the diminishes greatly, whereas it has been shown (e.g., Smith and Groat, 1979) that the frequency of errors on the word the is, if anything, increased when subjects are given explicit instructions to comprehend the material ,:hey are reading In any event, I plan to explore other l,%boratoryparadigms, including tachistoscopic identification and eye-movement monitoring, as well as the simple detection and proofreading tasks, in order to gather converging evidence for the processing units and rules that are implicated in the simple tasks. hmtually, I hope to construct and verify a model of the cognitive
Cognitive processes in reading text
125
processes involved in reading that will account for reading under a variety of conditions and explain, for example, why detection of letters is impaired by a word context under some conditions (e.g., Heaiy, 1976; Johnson, 1975) but facilitated by a word context under other conditions (e.g., Kn::y.er, 1970; Wheeler, 1970). If we can understand the processes that are responsible for reading efficiency, we will be in a position to tackle the important practical problem of finding ways to improve reading skill.
References Baron, J. (1978) The word-superiority eltect: Perceptual learning from reading. In W. K. Estes (ed.), Handbook of teurningundcognitive processes (Vol. 6). Hillsdale, NJ, Erlbaum, J 3 1.- 166. Corcoran, D. W. J. (1966) An acoustic factor in letter cancellation. Nature, 210, 658. Drewnowski, A. (1978) Detection errors ctn the word the: Evidence for the acquisition of reading levels. Mem. Cog., 6.403-409. Drewnowski, A. (1981) Missing Qrrg in reilding: Developmental changes in reading units. J exper. child Psychoi., 31, 154-168. Drewnowski, A., and Healy, A. r. (1977) I:.ttection errors on the and and: Evidence for reading units larger than the word. Mem. Cog., 5. 6X6-647. Drewnowski, A., and Healy, A. F. (1980) Missing -ing in reading: Letter detection errors on word endings. J. verb. Learn. verb. Behuv., 19, 147-262. Estes, W. K. (1975) The locus of inferential and perceptual processes in letter identification. J. exper. Psychol.: Gen.. 104. 122-145. Estes, W. K. (1977) On the interaction of perception and memory in reading. In D. LaBerge and S. J. Samuels teds.), Basic processes in reading: Perception and comprehension. Hillsdale, NJ, Erlbaum, l-25. Estes, ‘%‘.K., Bjork, E. L., and Skaar, E. (1974) Detection of single letters and letters in words with changing versus unchanging mask characters. Bul. Psychon. Sot., 3, 201-203. Garner, W. R., and Haun, F. (1978) Letter identification as a function of type of perceptual limitation and type of attribute. . exper. Psychal.: Hum. Percep. Pert, 4, 199-209. Gibson, E. J. (1965) Learning to read. Sci., ,‘48, 1066-1072. Gibson, E. J. (1971) Perceptual learning an*1 the theory of word perception. Cog. PsychoI.. 2, 351368. Cough, T. B. (1972) One second of reading. In J. F. Kavanagh and I. G. Mattingly teds.), language by cur and by eye: The relutionship btlrween speech und reading. Cambridge, Mass., MIT Press, 331..-358. Healy, .4. F. (1976) Detection errors on the word tht : Evidence for reading units larger than letters. J. exper. Psychol.: Hum. Percep. Perfi, .?, 235-242. Ht$y, A. F. (1980) Proofreading errors o:n the word the: New evidence on reading units. J. exper. Psychol.: Hum. Pemep. Per&,6,45-57. Healy, A. F. (In press) The effects of visual similarity on proofreading for misspellings. Mem. Cog. Healy, A. F., and Cutting, J. E. (1976) U:nits of speech perception: Phoneme and syllable. J. verb. Seurn. verb. Behav.. 15, 73-83. Healy, A. I’., and Levitt, A. (1980) Accessibility of the voicing distinction for learning phonological rules. Mem. Cog.. 8, 197-l 14.
W. (1949) The physkal principles of quantum theory (C. JMrart and C. Hoyt, trans.). %ork, Dower. (Originally published, 1930). pvchology and pedagogyof reudbzg.Cambridge, Mass., MIT Press. (Originally the function of letters in word identification: !Some data antl. a preliminary I. & R&. Learn. verb. Behav.. 14. 17-29. h time in a redundant visual display. J. exper. AychoZ.. 83, 391-399. rd a theory of automatic information protessing in read61975) Primary and secondary rebognition in reading. In D. W. Massan@csd.), Underlattguage: An information-prowsing analysis of speech perception,, reading, and isticr INewYork, Academic Press, 241-289. and Klitzke, D. (1977) Letters are functional in word identification. Mem. Cog., 5, Dig,
K. (1973) The perceptual reality of phonemes, sybbles, words, and sentences. vetb. Behav..,12.419-430. q, R. (1974) Gencc ef the word as a unit in the perception of language.
Toward an interactive model of reading. In S. Dornic (ed.), Httention and NJ, Erlbaum. ge, D. and Bremer, C. D. (1978) Units of word recognition: Eviden,ce for develges J. verb. Learn. verb. Behav., 17, 715-720. (1976)Decision times for wordand letter search: A who&tic word identification model A verb. Learn. verb. Behav.. IY. 93-11~1. Understandingreading: A p~cholinguistic analysisof reading and rearnlngto read. Vf. WiltMe,
Chat,
k 0979) Fwlling patterns, letter cancellation and the processing of text. In E. Wro!&rd and H. Bouma (eds.], Processingof visible&zngua~e (Vol. 1). New
e role of syllables in perceptual processing. &g. Aychol.. 5, of interference in serial verbal reactions. J. exper. Psychol.. 18, 643-662. rrsaro,D. W. (1973) Visual information and redundancy in reading. J. exper. 619711 Ikeoretka! analysisof an alphabetic confusion matrix. Percep. Psychophys., - (1978) Prowwe=in word recognition. &%g.Psych&, I, 59-85.
Cognition, 10 (1981) 127-134 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
127
nition in perception: Perceptualcoupling and unconsciousinfererm JULIAN HOCHBERG* CGhtnbia university
DO our perceptions of the world entail cognitive processes (comput,ations, inferences), or are they direct responses to ;he informative patterns of stimulation that act on our sense organs? To Helmholtz, as to J. S. Mill, channels of direct sensory response, which analyze the: patterns of stimulation into raw sensory responses, do not contribute directly to experience. The expectation of what patterns of sensation any given sensorimotor exploration would bring (the ‘permanent possibilities of sensation’) was taken as the percept. As Cassirer pointed out ( 1944), Helmholtz was saying that from the transformations in the pattern of sensations that result from exploratory movements, we perceive the invariances that are produced by, and specify, constant objects in the world. At least. three components must be distinguished : Stage A : The patterns of stimulation normally provided by the regularities in the physical world. Stage & : The channels of specific nerve energies. These result in preconscious events. They define the psychologically relevant effects of stimulation, but are not themselves psychologically accessible, and are known only by the constraints that they impose on what we can perceive; e.g., two points of light that stimulate onrly one and the same channel will be indistinguishable, so that discrimination psychophysics is one way to define the channels. Stage C: The patterns of expectations that have been learned from experience with A, above, viu B. These comprise mental structures (Hochberg, 1981), that reflect both the structure of the world, and the characteristics imposed by the sensory analysis in B. This last point amounts to what I have called Helmholtz’l Jle (Hochbl=rg, 1974, 1978): that we perlceive that object or event which would, under normal seeing conditions, be most likely to produce the pattern of sensations that we receive. To fit a perceived object or event to the sensary data in this way amounts to an unconscious inference, in which the premises are tb.e constraints in both A and B. *Reprint requests should be sent to Julian Hochberg, Psychology Department, Cofumbk University. New York, NY 10027, U.S.A.
128
Julian Hochberg
TO illustrate an inference based on A : isnFigure lA, let L = E x R, where E is illumination intensity, R is the object’s reflectance, and L is the light directed to the eye. The reflecta.:se, R, is a constant object property, whereas E, and consequently L, are continually changing. By taking information about E into account (i.e., the cues to illumination), the viewer might calculate or infer R, the object’s invariant surface color. Figure 1.
A. Of illumination (E) that an o&&t receives, it reflects some proportion (R) providing light (L,) to the eye L’, E’ and R’ are the hypothetical internal consequences of these. B. An object (target T) on its surround(S). C. If the upright object (U) is perceived as fast (F), it must also be perceived as receiving less illumination(E), in this arrangement.
A
B
In this cognitive theory of perception, one percept (E’, perceived illumination) causes another (R’, perceived reflectance). Such perceptual causation is difficult to d .,lonstrate c nvincingly,seems uneconomical,and the mentalistic use of a dependent variable as a causal independent variable makes many psychologists and physiologists acutely uneasy. Moreover, the notion of unwonscious inference, and the Helmholtz-Mill theory of perception, has been roundly criticized for more than a century. I will not address the most vocal of the critics, Gestalt theory, because I ‘believe that .both the figure-ground phenomenon, and the laws of organization, which together provide the heart of that theory, are explained far better in terms of the classical Helmholtzian theory (FIochberg, 1972, 19743, 1981) than by Gestalt theory itself. On the other hand, there are other lines of opposition, that cannot be so summarily dismissed, and that are highly active today. The first such attempt is simply to replace Helmholtz’s specific nerve energies with new ones that respond directly to aspects of the stimulus pattern which covary with the constant object characteristics. Hering ( 1878) and Mach ( 1886) proposed just such direct responses to a variety of object properties. As a very simple example of what this might mean: If the eye responded directly to ratios of adjacent intensities, LJL, in Figure lB, the sensory response R = 1*/L, will be proportional to the ratio of the reflectances, Rt/R,, assuming E equal for both. Versions of this proposal have been made repeatedly; it has some limited psychophysical support; and even has some counterparts
On cognition in perception
129
in physiological response (cf., Hochberg, 1972 for a review). It is plausible that such channels actually exist.Many other channels,each responsive to some ‘higher-order variable of stimulation’ (Gibson, 1950 j, have been proposed in recent years (Braddick et al., 1978; Graham, 198 1). Each such channel, if it exists, provides for the direct perception of some object property-in some limited sense of the word ‘perception’ -but psychologists concerned with perception must ask (i) whether that property can also be perceived in conditions that rule out the proposed channel, and (ii) how the information from such channels is combined where no one channel could possibly suffice. That is, such proposed new channels would shift the borders of Helmholtzian cognition around, but would not dispense with it. The direct theory of Gibson ( 1966, 1979) and his followers takes a much stronger position. To Gibson, we respond directly to the invariant relationships in the transformations of the pattern of stimulation, e.g., those that result from exploratory movements (cf., paragraph 1 of this Gaper). This approach reduces the task of perceptual psychology to physics, i.e., the analysis of the informative structure of the light at the eye (Stage A). Questions of information pickup (Stage B) are not addressed: if a potentially informative structure can be mathematically identified in stimulation, it is presumed capable of being used. Inference-like cognitive processes (Stage C) are occasionally admitted (Gibson, 195 1, 1979), but on4y when stimulus iuformatia3n is inadequate, and such situations are presumed (without evidence) to be rarely encountered. If cognitive processes do not Farticipate in perception when stimulus information is present, however, it is hard to see lhow they are acquired at ah. One wouid think that they must be learned in normally informed conditions. Are cognitive processes normalljl elicited? Note’ that under normal conditions, Helmholtzian and Gibsonian predictions rmst coincic:e (Hochberg, 198 1). Only in unusual circumstances, and in conditions (usual or unusual) in which stimulus information is incomplete, migbt we hope to find evidence of cognitive processes in perception, even were the:y in fact pervasive. One way to demonstrate the insufficiency of stimulrls information is to hold it constant whi.le varying the viewer’s perceptual premises; e.g., by Helmholtz’s rule, if one could change apparent illumination (E’) in Figure lA, then R.’should change accordingly, and there is a long history of observations that ju:t this occurs (cf., Hochberg, 1972, for a review). But how can we change E’ ezcept by making changes in the stimulrus distribution? If such a change were n;lide, the direct theorist could then attribute any perceptual effect to the stiml:lus c:iange. En one attempt to solve this problem (I;iochber& and Beck, 1954), three methods were used !o ch.ange a surface’s apparent orientation and, thereby, to change itI; apparent illumination: to monocular vision, U target in Figure 1C appeared to lie f!%t; tc’ binocular vision, it appeared up-
130
Julian Hochbtfg
right, and ‘therefore’ to receive less illumination, and ‘therefore’ to be of lower reflectance (since Lu remains unchanged). Because binocular vision might itself provide such lightness changes regardless of perceived orientation, orientat’ion was also changed by moving a white or a black rod behind it, or by moving the target from side to side by pulling on a wire, and in ah of these cases the target shifted from flat to upright appearance and changed apparent lightness accordingly. The binocular/monocular part of the experiment has been replicated, with similar results (Beck, 1965; Coren and Komoda, 1973; Gilchrist, 1977; Gogel and Mershon, 1969), more often than not (Epstein, 1961; Flock and Freedberg, 1970). Belatedly so.ne see this as strong evidence of unconscious inference (Epstein, 1977; Rock, 1977). But these experiments simply do not comprise strong evidence of unconscious inference, nor conclusive opp :sition to direct theory. Note that the replications all used the contracst between monocular and binocular vision to alter orientation (or to separate target and ground), and that represents a stimulus change En itself. Remember that if a change in stimulation results in a change in appearance, the latter may be directly caused by the former. Note too that apparent orientation might affect apparent lightness in ways that do not involve inference-like 3rocesses: Eye movements (and therefore successive cont:rast) may differ in the two conditions (Flock et al., 1966). Most important: subjects may compare different surfaces in arriving at their lightness judgments (Gogel and Mershon, 1969; Hochberg, 1972), and use different criteria for their judgments, wh.en comparing objects in different spatial layouts. That is, perceptual couplings (here, perceived lightness/perceived orientation) do not necessarily constitute evidence of unconscious inference. This means neither that the direct theory is invulnerable, nor that r ‘-::rg can servp *s strong evidence c f cognition in percept ion. We have long had s.rxing evidence that stimulus information is often simply insufficient to specify the very phenomena that the direct theory most confidently addresses, i.e., surfaces in motion at some slant to the viewer’s line of sight. The Ames trapezoid is correctly seen to rotate (the dotted arrow in F’igure 2B) when viewed from Figure 2.
A. Ames trapezoid, viewedfrom the front. B. Same, viewed from above. The true motion is shown by M; the perceived motion, by K’.
cl &A
p.
M
&NM, B
On cognitivein perception
131
quite close; from a few yards, however, it appears to oscillate as shown by the solid arrow (this is just what unconscious inference would yield if the sides’ convergence were taken for the depth cue of linear perspective, but we cannot accept that explanation without additional evidence). In this example, the information that specifies the true shape and motr)>n are present in the light at the eye; it is simply not effective. In many cases, much of the information is not present at all, because the object is part ially occluded ; t her1 elIen Gibson attributes the perception of the object, when based on the parts ildf occluded, to the operation of expectations, using language almost identical to that oi” Mill and Helmholtz. Furthermore, the information carried by fine texture and detail is above threshold only near the center of gaze (Hochberg, Green and Virostek, 1978). Information pickup then depends on the discrete glances that :he viewer elects to take. And because glances are not instantaneous, gre purposefully directed toward specific places, and are executed only to satisfy the viewer’s perceptual inquiry and then cease, the selection of the information picked up is itself dependent on schemas about the world, schemas that guide the glances, store the findings, and serve to sustain or terminate visunl inquiry. We do not know, therefore, whether sti ::ulus information is nomlally so accessible that perception could, even in principle, be direct. What little knowledge we have to that point (cf.,Hochberg, 198 1) is not encouraging. Surely in the areas of application that most need help from the perception psychologist - architecture, pictorial art (Hochberg, 19801, motion pictures (Hochberg and Brooks, 1978)-stimulus information must be inadequate, and cognitive processes must intervene in perception. Which means that c,uestions about the nature of mental structures are not idle ones. The notion that perception involves some sort of problem solving resurfaces periodically; most recently, spokesmen have been Gregory (1970), Epstein (1977)andRock(1977).Isharesomeofthisbelief(1968,1974a,b,1981)~but I think that most of these attempts fall far short of Helmholtz’s sophistication in that they make little or no attempt to consider the characteristics of Stage B in relating Stage A to Stage C. That is, although they are arrayed against Gibsonian direct theory, they in fact share its central feature inasmuch as they assume that -.lental structures merely mirror the world. That won’t, do. We know now that objects normally appear to have mutually consistent properties only because the objects themselves are physically selfconsislent, i.e., some of the fun&mental constraints of the physical world are not reflected in our mental structures. For exa.mple, although the viewer has the option of perceiving both Figures 3A and B as flat but consistent line patterns, only when the inconsistencies are in close proximity (cf.,Figure 3B)
132
Julian Hochberg
Figure 3.
A/B. Impossiblemires: Both are inconsistently oriented at 1 and 2, when the figure is viewed as a three dimensionalobject (note that linex must then change function along its length). Neither is inconsistent as flat line p&terns, but only B looks jlat. C Are ends I and 2 the same or opposite skies of the object’s surface? D. me orientation uf the cub2 is fixed at 1 so that the horizontal line must lie in front of the verticalline ar 2. After fixating point 2 for a few moments, however, reversalswilloccur.
L; A
Bl
B
8 B 2
D
does that occur (Hochberg, 1468). Such figures, discovered by Penrose and Pf,nrose (1958), are usually treeted as curiosities specific to line drawings but the phenomena that they reveal are in fad: far m.ore general: When real, three dimensional objects l&e that dr’awn in Figure 3C are shown to a subject under natural viewing cond.itions (i.e., binocular viewing, free head movements, etc.), he can tell whether cabs 1 and 2 are the same or opposite sides of the surface only by deliberately parsing the figure, even though aIl dihedrals are in full view (Klopfer and Hockberg, 198 1). As it was in Figure 3B, the relation between separated parts is not directly nor automatically evident. The failure lies in the perceptual process, not in the object nor in the stimulation it provides: because the viewer has no schema that will generate the relationship between parts, the stimulus information can only be used in a ‘bottom-up’ fashion. Another example shows us that schemas do not necessardy determine the appearance of their parts according to any principle of consistency or knplicity : although the drawing of a wire-and-cardboard+ object in Figure 38 is unambiguouslv; fixed at point 1, it reverses orientation when the gaze is directed for a few moments at point 2, and that is just as true for the real object itself as for the drawing (PetLrson and Hochberg, 198 1). Figure 30 re:l”utesall existing versions of Gestalt simplicity, including encoding theories I,,Attneave, 1954; Hochberg and MclUister, 1953 ; etc.). Figures 3 C 3nd 30 show us both that we must take mental structures into account in order to explain perception (and not merely Gibsonian stimulus informaLion), 3nd that those structures are not mere mirrors of the physical world. Even with objects be:fore our eyes, perceived structure is schematic, not dense and consistent as is physical structure.
On cognition in perception
133
For these reasons, it seems to me that none of the shortcuts tha: have been undertaken will work. We must study mental structure itself, its nature and origins, if we are tounderstand the cognitive processes at work in perception, and to assess their contribution. I ne various direct theories, and recent attempts to revive the view that perception is a cognitive process that draws upon mental structure, have either avoided Stage C entirely, or equated it with the constraints of the physical world. None of these shortcuts will work. All three stages must be studied conjointly, and we must study mental structure itself, if we are to understand the cognitive processes at work in perception, and to assess their contribution.
References Attneave, 1:. (1954) Some informational aspects of visual perception. Psychol. Rev., 61, 183-193. Beck, J. (1965) Apparent spatial position and the perception of Bghtness.J. exper. Psychor, 69, 17% 179. Braddick, O., Campbell, F. W., and Atkinson, J. (1978) Channels in vision: Basic aspects. In R. Held, H. W. Leibowitz and H. L. Teubea (eds.), Handbook ofSensou Physiology, Vol. 8. New York, Spriuger, pp. 3-38. Cassirer, E. (1944) The concept of group and the theory of perception.Philos. Phenom. Res., 5, l-35. Coren, S. C., and Komoda, M. K. (1973) The effect of cues to illumination on apparent brightness. Am, J. PsychoI., 86. 345-349. Epstein, W.. (1961) Phenomenal orient&an and perceived achromatic color. J. Pqvchol., 5.2, 51-53. Epstein, W, (ed.) (1977) Stab%@ and Conrtmcy in Visual Perception: Mechanisms and Processes. New York, Wiley. Flock, H., and Freedberg, E. (1970) Perceived augle of incidence and achromatic surface cola-. Percep. Psychophys., 8, 251-256. Flock, H., Wilson, A., and Poizner, S. (1966) Lightness matching for dsferent routes through a compound scene. Percep. Pqvchophys.. I, 382-384. Gibson, J. J. (1950) 77sePerception of the VisualWrZd. Boston, Houghton MiffIin. Gibson, J. J. (1951) What is form? P@&or Rev., 58,403-412. Gibson, J. J. (i966) The Senw Considered as PerceptualSystems. Boston, Houghton Mifflin. Gibson, J. J. (1979) The Eco&iculApproach to VisualPerception. Boston, Houghton Mifflin. Gilchrist, A.. (1977) Perceived lightness depends on perceived spatial arrangement. SC& 195,185 -187. Gogcl, W., and Mershon, D. H. (1969) Depth adjacency in simultaneous contrast. Percep. Psychophys., 5,13-17. Graham, N. (1981) Spatiaifrequelwy channelsin human vision: Detecting edges without edge detectors. In C. Harris (ed.), VBwaJCBdi@md A&ptubility. Hillsdale, NJ, Erlbaum. Gregory, R. L, (1970) l%e Inte&ent me. London, Weidenfeld. Hering. E. (1964) Outlines a Theory oftire Light Sense. Translated from German by L. M. Hurvich and :D. Jameson, Cambridge, Mass., Harvard University Press. (Origitiy wblished 1878). Hochherg, J. (1968) In the mind’s eye. In R. N. Haber (ed.), Contempomfy new andRe@?arch in V&u.?Pemeption. New York, HoEt, Rinehart and Winston.
of
134
Julian Hixhberg
Hochberg, J. (1972) Perception, 1. Color and shape. II. Space and movement. In J. W. Kiing and L. A. Riggs (eds_), wooelworth u&J schz or be&s Experimental Psycho&y. Third edition. New York, Halt. Rinehart and Winston. Hochberg, J. (1974a) Higherorder stimuli and interresponse coupling in the perception of the visual w&d. In R. B. Macleod and II. L. Pick (eds.), Perception: Essaysin Honor of James J. Gibson Iti~ca, Cornell University Press, pp. 17-39. Hochbelg, J. (1974&) Organization and the Gestalt tradition. In E. C. Carterette md M. Friedman; (eds.), Handbook of Perception, Vol. 1. New York, Academic Press. Ho&berg, J. (1978) Perception Second edition. EngIewood cliffs, NJ, Prentice-Hall. Hochberg, J. (1980) Pictorial functions and perceptual structures. In M. Hagen (ea.), The Perception of Pictures, Vol 2. New York, Academic Press. Hochlberg, J. (1981) Levels of perceptual organization. In M. Kubovy and J. Pomerantz (eds.), F’erceptual Orgmizatbn Hillsdale,NJ, Erlbaum. Hochberg, J.,and Beck, J. (1954) Apparent spatial arrangement and perceived brightness. J. exper. Psy chd, 47, X3-266. Ho&berg, J., and Brooks, V. (1978) The perception of motion pictures. In E. C. Carterette and M. Friedman (eds.), Handbook of Perception, Vol. 10. New York, Academic Press. Hochberg, J., Green, J., and Virostek, S. (1978) Texture conchrsion requires centrai viewing: Demonstrations, data and theoretical implications. Unpubhshed paper delivered at the APA Convention, Toronto. Ho&berg, J., and McAIister, E. (1953) A quantitative approach to fgural ‘goodness’. J. exper. Psychol. 46.36 l-364. KIcpfer, D., and Hochberg, J. (1981) Seeing is not perceiving: Schemas are needed even when visual inforrxtion is complete. Proc. EasternPsychobgikal AsmcWm, 148 (Abstract). Mach, E. (1959) 7%e Analysis of Sensations. Translated from the 5th German edition, 1886, by S. WaterIow. New York, Dover. Penrose, L., and Penrose, R. (1958) Impossible objects: A special type of visual illusion. Bra?..I. P~JCM. 49,31-33. Peterson, M. A., and Hochberg, J. (1981) Perspective reversals that refute both Gestalt and Direct th.zories of objects perception: Measures of local cue strength and attention. .rRoc. Eastem Psychologicd Assocktbn, 148 (Abstract). Rock, I. (1977) In defense of unconscious inference. In W. Epstein (ed.), Stub&y and Constancyin Visrrol Pereeptidn: iUechanismsand Proceszx New York, Wiiey.
cognirion, 10 (1981) 135-137 @Elsevier Sequoia &A., Lauisanne- Printed in The Netherlands
135
Specialized channels for cognitive responses M. JEANNEROD” Laboratoire de Neurops ychologie Exptkimen ta/e
In this note, I will try to show that there are several modes of response to external events,’ and that cognitive responses, from which cognition is inferred, is one. Also, I will advance the hypothesis that responses to external events are mediated by separate neural channels each of which is characterized by a given inputautput relationship and thence by a given production. As an example, in visuomotor behavior the different properties of visual objects (e.g., shape, color, spatial location) are not processed by the same neural structures. They are matched by specific mechanisms which generate motor commands appropriate for each property (for a more complete version of the hypothesis, see Jeannerod, 198 1). Now the question is this: 2re there cognitive chat,nels? By this I mean do channels exist that deal with the cognitive aspect of a situatiol.1 and produce cognitive responses? A visual object for instance may be subjected to different levels of processing, besides that effected by visuomotor channels. Proerlction of a verbal response, mental representation, compL -ison with other objects, etc., are based on the same object properties as tho *efeeding into visuomotor structures. Can they be conceived as resulting from the activity of other parallel, specialized neural mechanisms? My arguments are drawn frohm the study of brain-lesioned subjects. Over the last ‘ntidred years neurological observation has accumulated a considerable amount ol information that seems to have been largely neglected by people involved Jn cognitive science. It is good news to hear that they have finally dkovered this ignored continent (Posner et al., 198 1). Subjects with bilateral lesions of the visual cortex of the brain may displry the typical syndro.ne of cortical blindness, once called ‘psychic’ blindness. They do not see, b,y the conventional meaning of the word ‘see’ since they do not experience visual changes or events. Some of them may even be unaware of not seeing, and may report hallucinatory visual scenes unrelated to the on-going physical reality. Yet, in cortical blindness, only part of the visual system is actually altered by the lesion. Other pathways also originating in the retina but iterminating in areas outside the visual cortex are anatomi*Reprint requem should be addressed to M. Jeannerod, Laboratoire de Ncuropsychologie Exp&i-
mentale,INSERM LJ94,69500 BIOB,France.
136
M. Jannemd
afly intact. As a matter of fact, visual responses such as eye movements to bright moving objects can still be elicited from these patients, although they remain unaware of the stimulus. Thus, in the case of psychic blindness only a certain mode of neural processing can be effected up to the point where it leads to an appropriate response:. This response reflects the activity of
surviving structures. The cortical lesion cuts off the access of visual information to the ‘mental organ’ specialized for cognitive operations, although the same information remains available to the visuomotor organ. According to the channel hypothesis the visuomotor organ only does what it is built for, that is, it extracts a limited number of parameters from the visual world and produces the corresponding responses. A more dramatic illustration of this point is given by another group of patients with more localized lesions of the visual cortex. In the case of unilateral lesions? cortical blindness is limited to a particular area of the visual field (the scotoma) in the half field contralateral to the lesion. It has recently been observed that the scotoma is only a ‘relatively’ blind area, from where responses to visual stimulation can still be obtained. The important point is that whether a response can be obtained or not depends on the requirements of the task. If the task requires a verbal response based on subjective experience of the stimulus (using the common perceptual or cognitive ZX&), no response is given. If, on the other hand, the subject is forced to 3each’ (by eye or by hand), fcr a stimulus briefly presented within the scotoma a clear visuomotor response can be recorded. Since the subject remains unaware of the stimulus, he or she experiences guessing rather than seeing. This is true not only for spatial location uf stiiuIi but also -for more intricate properties like shape or size (in this case the forced choice procedure is also used, the subject being asked to show on cards which of two patterns has been flashed within his scotoma). For details see Berenin and Jeannerod (1979). In other words, these patients are able to locate visually objects which they duo not see. Furthermore, they may eventually be able to detect intrinsic properties in objects of which they are totaIIy unaware. This is an example of the functi;Ang of discrete neural mechanisms subserving discrete behavioral productions. The lesion artificially @its behavior into modular compartlments and isolate: segmental operations which normally cannot be dissociated. At this point, the relevance of the concept of ‘mental organs’ (Chomsky, 1980) and of ‘cognitive channels’ to cognitive functioning, has to be dirscussed.These two terms have been taken here as equivalent. I consider the channel as a distributed neural ensembIe bearing a particular transfer function, i.e., rt?ceivinginformation from the external world or from other parts of the brain and releasing a predictable production. By this definiti In mental
Specidized channels for cognitive responses
137
production is not considered as autonomous with respect to other produclions of neural activity. It is only the input-output relationships which determine the specificity of a given channel, not the internal arrangement of its connections, which can be suspected to be similar to those of any other channel. In other words, the channel concept would definitely depart from current conceptions of mental operators elaborating their production from the inside, du: to .A particular ‘genius’ of their internal structure (for example, see Popper :bnd Eccles, 1977). The argument underlying this discussion is that attributir;g to the mind structural properties which are not operational for other aspects of behavior will unavoidably result in a revival of the concept of autonomy of the ‘mental level’. The mental-level concept implies the existence of some undefined ‘higher level’ of brain activity subserving cognitive ‘functions’. Such an entity, however, has only a small chance of finding a precise embodiment in neurological terms, except that of a crude ‘localization’ in a given brain area. The notioq of a mental level only concedes that cognition may have ‘something to do’ with neural mechanisms and that once a certain degree of complexity of a neural ensemble has been attained a new set of properties could emerge from that ensemble. In other words, although the mental level would admittedly be brain-dependent, it would not be reduced to the sum of simple neuronal operations like those which are currently described for simpler lerrels of activity. The problem hers! is not with the lack of a satisfactory description of neuronal operations or connections which could account for cognitive responses (though this may be a real problem): it is rather with the incompatibility of the levels of explanation neural functioning on one hand and for mental functioning postulated ? on the other. In this context, neurai mechanisms would represent but a useless ornament for cognition, and the whole concept of mental level would fail to reach a heuristic value, due to its dualistic and teleolo@cal underlying nature .
References Chomsky, N. (1980) Rules and representations. Beku. Bruin LG., 3. I-63. Jeannerod, M. (1981) Intersegmental coordination during reaching at natural visual objects. and A. Baddeley (eds.), Attention and Performunce IA’. Hillsdale, NJ, Erlbaum, pp Perenin, M. T., and Jeannerod, M. (1979) Subcortical vision in man. Trends Akurosci., 2, Popper, K., and Eccles, J. C. (1977) The LSeffond its Bmin. Berlin, Springer International. Posner, M. I., Pea, R., and Volpe, EL(In press) Cognitive Neumscience, Toward ti Science of
In J. Long 153-168. 204-207.
Synthesis.
CogniPion, 10 (1981) 139-143 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
139
Cognition, computers, a@ mental models P. N. JOHNSON-LAIRD” Lhiversity of Sussex
Two qilestions ought to haunt any student of cognition. First, is it possible to achieve a scientific understanding of the mind? It is to be, hoped, of ccurse, that a complete science is impossible since it would probably destroy the consciousness of free will. Second, are there profound uniformities in the *~ys in which t.he mind works? For example, if we understood how speech is perceived, would we thereby advance our understanding of, say, the visual perception of shapes ? These two questions are presumably related in that a positive answer to the second is likely to lead to a positive answer to the first. .;y own research certainly inclines me towards supposing that indeed tfiere are underlying l_miformities in thought. For many years, I worked alternately on reasoning and comprehension. If I became stuck in one area, then I would switch to the other. Unfortunately, this strategy his recently been denied to me by the discovery of an underlying communality in the two areas. It concerns the role of mi=ntal models. My first inkling that mental models might be important in comprehension is reflected in the following: It is possible that from the meanings of sentences in a connected discourse, the listener implicitly sectsup a much abbreviated and not especitily linguistic model of the narrative, and th? recall is very much an active reconsnructionbased on what remains of this model, Wherethe model is incomplete, matAal may everibe unwittingly invented to render the memory more meaningful or more plausible-a process 4rvhichhas its parallel in the initial construction of the model. A good writer or raconteur perhaps has the power to initiate a process very similar to tb.e one that occurs when we are actually perceiving (or imagining) events instead of merely reading (Johnson-L&d, 1970). or hearing about them.
Bransford and his colleagues, of course, advanced a similar ‘constructive’ theory of comprehension and gathered convincing evidence in its support (see e.g., Bransford and McCarrell, 197%). My second inkling about mental models arose from studying syllogisms deductive inferences of (Wason and Johnson-Laird, 1972). These a*rilv +ple 011.
*Reprk,rtrequests should be sent to P. N. Johnson&&d, Centre for Research on Perception and Cognition, I&oratory of Experimental Psychology, University of Sussex, Brighton, BNJ9QG, England.
f40
P. N. hhnson-Laird
the form
:
San:, of the scientists are parents All of the parents are drivers Some of the scientists are drivers Syllogismt itre a mce test case for the feasibility of a cognitive science. There are only 64 possible forms for their premises, and if we are ever to understand anything about mental processes, we ought to be able to understand how people draw conclusions from them, They were first studied experimentally at the turn of the century, yet we still have no complete understanding of how human beings cope with them. ,4 seductive hypothesis is that there is some sort of mental logic, perhaps based on representations akin to Euler circles; for a time, i certainly subscribed to such a doctrine. A major problem with it, however, is that it gives 1:~ very ready account of either the ‘figural’ effectsubjests tend to draw conclusions like the one illustrated in the example above rather than itsequdly valid converse -or the systematic errors that they make (see Johnson-Laird and Steedman, 1978). How do people mentally represent syllogistic premises? No psychologist was ever able to tell me, but several subjects reported that they formed images of the states of affairs described in premzses. The pattern of errors that one observes is certainly compatible with the idea that subjects in general construct mental models of the premises, whether or not they take the form of images. For example, suppose you present your subjects with the task of drawing a conclusion from the following premises: All of the artists are beekeepers All of the chemists are beekeepers They can form a mental model of the first premise by imagining an arbitrary number of artists and identifying each of them as a beekeeper. Since there may be beekeepers who are not artists, they, too, must be represented in the model. Its structure must accordingly take the following form:
a=b a=b a=b fb) 0) Ib) where each ‘a’ represents an artist, each ‘b’ represents a beekeeper, and the parentheses indicate that an individual may, or may not, exist. In adding the information from thti second premise to this model, a logically prudent subject should consider all the different ways in which it can be combined. Some subjects evidently consider only this combination:
Cognition,corqxdters,oiidmentalmodels
1141
a=b=c a=uu=c a=b=c (b) (b) (b) They draw the invalid conclusion that aZlof the artists are chemists, or its invalid converse, al! of the chendsts are artists. Other subjects also consider the combination: a=b=c a=b=c a=b b=c (b) (b) They refrain from the previous conclusions, but draw the equally invalid conclusion that s:,)rre of the artists are chemists, or its i.nvalid converse, some bf the chemists are artists. Fortunately for the rational reputation of the human race, about half the subjects that we have tested evidently consider $he further combination : a=b a=b a=b b=c b=c b=c They correctly reply that there is no valid conclusion interrelating the artists and chemists. The figural effect remains something of a mystery. It could refleet an inherent directional bias in the structure of mental models, or alternatively, the process of forming an integrated mtidel in working memory. My current conception of comprehension is that there i.s an initial rapid translation of an utterance into itssuperficial linguistic form, followed by an optional process in which this representation is used in the construction of a mental model. The process of model building goes beyond the literal content of th.e utterance since it relies on inferences based1 on general and specific knowledge. It also rehes on a ‘procedural semantics’ (see Miller and JohnsonLairdl, 1976; Johnson-Laird, 1977) in which the meanings of words play an appropriate part in constructing models ab inndtio, adding information to them from subsequent utterances, verifying senterxes with respect to them, i;7ud recursively manipulating them in order to check whether there is any way in which a sentence could be consistent (or inconsistent) with the prior discourse.
142
P. IV.Johnson-Luiid
The claim that sentences can be verified with respect to mental models should not be taken as a species of ‘verificationism’: it is one thing to compare a sentence with a mental model, quite another to verify it in reality. Most of these processes occur in inference, too. Indeed, the logical properties of mauy words emerge directly from the semantics that is required to construct models from them. There is no need to postulate rules of inference governing their behaviour. There is no need to postulate a mental logic: the same principle applies equally to sentential connectives such as and and or, and, as we have seen, to quantifiers such as some and all. The advantage of this approach is that it solves at a stroke the problems of which particular logic or logics are in the mind, how they are mentally specified, and how children acquire them. These issues are cut off without a source, because logic is banished from the ,mind. When one considers the follies and horrors of the {luman predicament, it may be tempting to suppose that rationality is thereby banished, too. However, the fundament:*ll semantic principle governing both the truth of a general assertion and the validity of an inference is that there should be no counterexamples. Some people at least are aware of this principle, and the experimental evidence suggests that they search, in a more or less haphazard way, for models of premises that are inconsistent with the putative conclusions that they have drawn. It is important to emphasize that the search appears to be neither systematic nor exhaustive, beca.use the absence of these characteristics is the best evidence we have that deductive thinking is not guided by mental logic. The the&es that my colleagues and I have developed in order to account for these phenomena have often been modelled in the form of computer programs. There are obvious analogies between the operations of the mind and the execution of a computer program- a relation that was not lost on Kenneth Craik (1943), who was the first psychologist to suggest that reasoning might consist of the manipulation of models of reality, and he was writing several years before the invention ofthe programmable digital computer. However, there is another more important reason for computer modelling. Theoretical intuitions are very valuable (to those that ‘*havethem) but, if they are needed to work out what a theory predicts, there is a strong possibility that they are responsible for the predictions, Fnd that the theory itself has no explan;,ltory value. It is not a signpost, but a crutch on which the theorist lean:; in order to point the way. A simple criterion that avoids this danger is to check that those components of the theory that give rise to predictions are describable in the form of an effective procedure, i.e. they can be ipressed in the form of a working computer program. This criterion does not imply that the mind is nothing but a computer. It may turn out that the mind uses functions that are not computable- it is easy enough to prove the existence
Cognition, computers, and mental models
143
of non-computable functions. But this phenomenon would, of course, place a strong limitation on the possibility of a scientific psychology. Likewise, it may turn out that someone will succeed in refuting the thesis that all effective procedures are computable. At present, however, any scientific theory of the mind should certainly be restricted to an effective procedure. To abandon this criterion is to allow that theories CY~? be vague, confused, and, like mystical doctrines, only properly understood by their proponents. Nevertheless, although 11have sketched an optimistic answer to my initial question about possible uniformities in mental processes, the answer to my first question may be negative: there may be certain aspects of human mentality that cannot be captured in any theory that tag be modelled by a computer program. References Bransford, J. D. and McCarreII, N. S. (1975) A sketch of a cognitive approach to comprehension: some thoughts about what it means to comprehend. In W. B. Weimar and D. S. Palermo (eds,), %gnirion and the Symbolic Processes. Hi&dale, NJ, Erlbaum. Craik, K. (1943) 77re Nature of Explanation. Cambridge, Cambridge University Press. Johnson-Laird, P. N. (1970) The perception and memory of sentences. In J. Lyons (ed.), New Horizons in Linguistics. Harmondswarth, Middx., Penguin. Johnson-Lain&P. N. (1977) Procedural semantics. Cog, 7,189-214. Johnson-Laird, P.N. and §teedman, M. J. (1978) The psychology of sylIogisms. Cog. Psychd, IO, 64-99. Miller, G. A. and Johnson-Laird, P. N’. (1976) Language and Perception. Cambridge, Cambridge University Press; Cambridge, Mass., Harvard University Press. Wason, P. C. and Johnson-Laird, P. N. (1972) The Psychology of Reasoning. London, Batsford.
Cognition, 10 (1981) 145-150 @Elsevier Sequoia S.A., Lausanne - Printed in ‘TheNetherlands
145
Capturiq
attention
JOHN JONIDES* DAVID
E. IRWIN
University of Michigan
The first chapter of WilheLn Wundt’s text, An Introduction to Psychokogy, is devoted to the topic of attention. This reflects attention’s prominent role in the history of investigatio:ls of cognition and perception. And deservedly so. Humans and other animals are limited processors of information. l%ctiuse of this, a proper understanding of tl\e inner workings of mental mechanisms that transform and diges,t information must include a description ol the processes by which certain sources of input are selected fcr further analysis while others are ignored. Although research on the topic of si:iective attention has been fairly eciectic in its choice of paradigms and specific phenomena, recent research has concentrated especially on two issues. The first concerns spatial allocation of processing resources, and the second tocusses on details of processing when there is little limit on our capacity to engage in mental activity. Below we briefly review developments in each of these areas and tentatively offer some i promoses for the near future.
Spatial selectivity In 19 12, Wundt commented: ‘If . . we practice letting our attention wander over . . .different pa.& of the field of vision while keeping the same fixationpoint, it will soon be clear to us that the fixation-point of attention and the fixation-point of the field of vision are by no means identical’ (p. 20). This early work coupled rith ihe research of Purkinje and Helmhoitz on related issues has, from time to time in the history of perceptual research, spurred psychoiogi&s to inquire about the processes involved in attending to spatial locations. Speriing’s (1960) research with the partial report techniqde and Cherry% (1953) studies of dichotic listening can, perhaps, be pinpointed as the developments that have renewed concern with this problem. *Reprint requests should ix sent to J. Jonides, hychology Depaatmed, University of Michigan, Human Performance Center, 330 &ukar~Rd., Ann Arbor, Mich. 48104, U.S.A.
146
John Janides and David E. Irwin
In the ensuing years, research by various investigators has uncovered some of the details of the processes involved in selectively attending to spatially defined stimuli (e.g., Eriksen and Hoffman, 1973; Hoffman, 1979; Jonides, in press; Posner, 1980). One of the most important findings is that spatial sele.cc::ity for visual as well as auditory stimuli can be accomplished without any overt change in the peripheral sense organs (Eriksen and Hoffman, 1973; Jonides, 1980; Pasner, Nissen and Ogden, 1978). This fact has led to a debate about thi: locus of selectivity in the processing stream, a debate that in many ways mimics the debate concerning the adequacy of selective attention models tha.r have been developed since Broadbent’s (1958) seminal work in this area. For example, Shiffrin and his colleagues (Shiffrin and Gardner, 1972; Shiffrin, McKay and Shaffer, 1976) have argued that selection occurs in short-term memory after early perceptual analysis has been completed.. The experiments of others, hc’wever, indicate a selectivity thah is difficult to reconcile with a memory interpretation, since the tasks in which the selectivity occurs place only trivial memory demands on subjects (e.g., Bashrinki and Bacharach, 1980; Jonides and Somers, Reference note 11; Posner et aZ., 1978; Shaw and Shaw, 1977). Resolution of this conflict requires further investigation, and a coherent synthesis of a growing body of research. One essential component of a theory of selectivity will be a model of the actual mechanism of selection, regardless of its locus in the processing stream. While some of the papers cited above allude to such a mechanism, to date there has been insufficient attention to this problem. Shaw and Shaw (1977) proposed a general moclel of selectivity that has the important feature of being general across specific tasks. Jonides (1980) has tested specific versions of this model that seem to narrow the space of remaining alternatives to an interesting subset-namely, those in which processing occurs in parallel over a variety of spatial loci and can be focussed on one of these loci by internal guidance or by external stimuhrs control (Jonides, in press). But it is clear that more specific models need to be proposed and tested before progress can be made on this problem. Once we have a better understanding of the mechanics underlying spatial seiectivity in vision, the reiationship between shifts of attention to local spatial regions and shifts of the eyes to spatial locations may be uncovered. Interesting parallels between these phenomenn suggest that they may share some fundamental mechanisms in ccmmon (see e.g., Jonides, in press, and Todd and Van Gelder, 1979, for disc?.:qzions of internal versus external control over the body’s and the mind’s eye movements). Although eye movements seem to be neither necessary (Eriksen and Hoffman, 1973; Jonides, 1980; Posner ef al., 1978) nor sufficient (Klein, 1980; Remington, 1980)
Ca.qturingattention
147
conditions :for attention movements, the question of whether movements of attention can facilitate the programming or execution of subsequent eye movements has received less study (Todd and Van Gelder, 1979). Neurophysiological research by Wurtz and colleagues (Goldberg and VWurtz, 1972; Mohler and Wurtz, 1976; Wurtz and Mohler, 1976) suggests a further basis for pursuing the connection between attention shifts and eye movements. Their work has suggested that cells in the superficial layers of the superior colliculus are involved in the control of both eye and attention movements. Several other investigators, however, have found evidence that cells in the parietal lobe are involved in shifts of spatial attention, independent of eye movements (e.g., Robinson, Goldberg and Stanton, 1978). Further research in this area has important implications for models of se!ective attention and saccade guidance and control (e.g., Mays and Sparks, 1980). Automaticity In a sense, a second recent Zicus of attention research has not been on attention at all, but ra:her its absence. The topic is nicely introduced by a quote from William James’ Principles of Psychology : ‘If an act became no easier after being done several times, if the careful 3rec”ion of consciousness were necessary to its accomplishment on each occasion, it is evident that the whcle activity of a lifetime might be confined to one or two deeds -that no progress could take place in development. A person might be occupied all day in dressing and undrc~sing himself; the attitude of his body would absorb all his attention and energy; the washing of his hands or the fastening of a button would be as difficult to him on each occasion as to the child on its first trial; and he would, furthermore, be completely exhausted by his exertions. For :rrhile automatic acts are accomplished with comparatively little weariness, the conscious effort of the will soon produces exhaustion.’
During the past several years, psychologists have rediscovered this distinction between processes that are under strategic control and those that are automatic, especially in the domain of perceptual and cognitive tasks. One might argue that strategically controlled processes typically constitute our most impressive armament against complex problems, and that they stand in contrast to the more stereotyped activities that have been identified as automa.tic. Even in their relative stereotypy, however, automatic processes are not a mere cedilla in our mental lives, as James indicated: They form a crucial part of our processing repertoire becauc- they require little effort and attention to execute. Consequently, we are left free to devote our sophisticated mental machinery to the ttasks that require it.
148
John Janides ard David E. Inuirt
These considerations have led psychologists in recent years to concentrate on the development of automaticity in processing. The highlights of this work can be found in LaBerge (1975), Logan (1978), Schneider and Shiffrin (1977), and Shiffrin and Schneider (1977) among other places. The empirical work in these yapers tries to identify the Iearning regimens that are necessnry and/or sufficient for automaticity to develop. This is obviously one of the crucial issues that must be addressed in research on automaticity, and so it is not surprising that the vast majority of work on this topic sin;;e Bryan and Harter’s (1899) early study has concentrated on developmen”,. But some might argue that this emphasis has been premature because the successfui study of automaticity first requires a wellspecified, theoretically-motivated set of criteria that can be used to identify when a process has become automated. Only a few investigators have tried to estaalish such criteria with any empirical tests of their adequacy (see Jonides, in press; Jonides, note 2; Logan, 1978; Regan, 198 1). Consequently, it seems reasonable to prescribe a substantial emphasis on this problem before further work on the d.evelopment of automaticity proceeds apace. The benefit of a well-defined set of empirical criteria will extend beyond the study of development to another important aspect of automaticity, its generality. As is reasonable, most of the available research on automatic processing has demonstrated its development within a single task corrtext, or at best witnin the context of two tasks that are very closely related (e.g., Schneider and Shiffrin, 1977). While this is an important first step, it leads one to ask about the extent to which an automatic process once developed will transfer to a new task situation. Clearly, automaticity would lose much of its current play if there were convincing evidence that automatic processzs are completely task specific. Of course, it will not be straightforward to test the generality of automatic processes since cognitive psychology does not yet have a taxonomy of processes that participate in various t;ask performances. Nevertheless, this problem should attract some empirical attention over the coming years, as it has already begun to do (Benjamin and Jonicfes, Reference note 3). We can identify one final theoretical issue about autabmaticity that is worth noting as weli. There seems to be an undercurrent of belief in the literature that automatic and non-gutomatic processes are not merely ends of a continuum, but are qualitatively different from one anoihler. The case for this belief has not been made, however. Indeed, judging from the course of previous theoretical arguments about incremental versus all-or-none leaning, it will not be a case that is easy to make. Nevertheless, it iis a fundamental question whose answer will help determine the form of specific models of automatic processes that are proposed.
Capturing attention
149
Summary
The topics of spatial selectivity and automaticity have formed the focus of much rece,It research on attention, both in our laboratory and elsewhere. They do not, of course, nearly exhaust the possible areas for further research. For example, Triesman’s (Triesman and Gelade, 1980) research on the application of ‘focal attention’ to encodi;;g is certain to excite interest, especially because it seems to contradict much of the recent work on automaticity.in encoding (e.g., Egeth, 1977). Alsc, there is a growing concern that phenomena investigated ir the experimental laboratory have direct application in the ‘real world’. This suggests that advances in attention research may lead to advances in the diagnosis and treatment of a range of pathological conditions that may be due in part to attentional deficits, including autism, schizophrenia, and various kinds of brain trauma. Whatever these additional developments, to reach an understanding of the mechanisms underlying spatial selectivity and automaticity would be an important theoretical breakthrough in the decade ahead. References Baslcinski, H. S. and Bacharach, V. R. (1980) Enhdncement of perceptual sensitivity as the result of selectively attending to spatial locations. Percep. Psychophys., 28, 241-248. Broadbent, D. E. (1958) Perception and communication. London, Pergamon Press. Bryaln, W. L. and Harter, N. (18991 Studies on the telegraphic language: the acquisition of a hierarchy of habits. Psychoi. Rev., 6, 345-375. Cherry, E. C. (1953) Some experiments XI the recognition of speech with one and two ears. .I. acousf. Sot. Amer., 25,975-979. Egeth, H. (1977) Attention and preattention. In G. H. Bower (ed.), The psychology offearning und motivafion, Vol. 11. New York, Academic Press. Eridksen, C. W. and Hoffman, J. E. (1973) The extent of processing of noise elements during selective encoding from visual displays. Percep. Psychophys., 14, 155-160. Goldiberg, M. F. and Wurtz, R. (1972) Activity of superior colliculus in behaving monkey: Effect of attention on neuronal responses. J. Neurophysiol., 3.5, 560-574. Hoffman, J. E. (1979) A two-stage model r)f visual search. Percep. Psychophys., 2.5, 319-327. James, W. (1890) The principles ofpsychcb’>gy. New York, Henry Holt. Jonides, J. (1980) Toward a model of the mind’s eye’s movement. &n, J. Psychok, 34, IO?-1 12. Jonides, J. (In press) Voluntary versus automatic control over the mind’s eye’s movement. In Long, J. B. and Baddeley, A. D. (eds.), Attention and Performance IX. Hillsdale, NJ, Lawrence Erlbaum Associates. Klein, R. (1980) Does oculomotor readiness mediate cognitive control of visual attention? In R. Nickerson (ed.), Attention and Performance VZZI.Hillsdale, NJ, Lawrence Erlbaum Associates. LaBerge, D. (1975) Acquisition of automatic processing in perceptual and associative learning. In P. M. A. Rabbitt and S. Dornic (eds.), Atttintion and Perjbrmance K London, Academic Press.
150
John Jonides and David E. hwin
bgan, G. D. (1978) Attention in characterclassification tasks: Evidence for the automaticity of compone nt stages. J. exper. Psychol.: Gen., IO 7, 3 2-63. Map, L_ E. and Sparks, D. L. (1980) Dissociation of visual and saccade-related responses in superior colliculus neurons. J. Neurophys... 43, 207-232. Mohler, C W. and Wurtz, 12. (1976) Organization of monkey superior colliculus: Intermediate layer cells discharging before eye movements. J. Neuroph~s., 39, 722- 744. Posner, M. I. (1980) The orienting of abteatbn. Q, J. exclcr. Psychol., 32, 3-25. Posner, M. li., Nissen, M. J., and Ogden, W. C. (1978) Attended and unattended processing modes: The role of set for spatial locaeioil. In H. L. Pick, Jr., aisd E, Saltzman (eds.), Modes ofperceiving undptvcessing information. Hillsdale, NJ, Lawrence Erlbaum Associates. Regan, J. E. (1981) Automaticity and learning: Effects of familiarity on naming letters. J. ?xper. Psychol.: Hum. Percep. Perf0.m.. 7, 180-I 95. Remington, R. W. (19801 Attention and saccadic eye movements. J, exper. Psychol.: Hum. Percep. Perform., 6,726-M4.
RobMon, D. L.,Goldberg, M. E.,and Stanton, G. B. (1978) Parietal association cortex in the primate: Sensory m~:hacism~sand behavioral modulations. J. Neurophys., 41, 910-333. Schneider, W. and ShifZrin, R. M. (1977) Controlled and automatic human infr~~matbn processing: I. Detection, search, and attention. Psychol. Rev., 84, l-66. Shaw, M. L. and Shaw, P. (1977) Optimal allocation of cognitive resources to spatial location. J. exper. Ps:dol.:
HtFrn.Percep. Perform., 3. 201-211.
Shiffrin, R. and Gardner, G. (1972) Visual processing capacity and attentional control. J. exper. &y&al.. 9& 72-83. Shiffrin, R. ‘M.,McKay, D. P., and Shaffer, W. 0. (1976) Attending to 49 spatial positions at once. J. erper. P~yc~ol.: Hum. Pewep. Perform., 6. 190-215.
Shiffrin, R. M. arid Schneider, W. (1977) Controlled and automatic human information processing: II. Per;ept:;aJ learning, automatic attending, and a general theory. P$ychol. Rev., 84, 127-190. @e&g, G. (196(1~)Theinformation available in brief visual presentations. Psychol. Mono., 74, No. 11. Todd, J. T. and van Gelder, P. (1979) Implicaations of a transient-sustained dichotomy for the measurement of human performance. J. exper Psychol.: Flum. Percep. Perform., 5, 625-638. Triesman, A. %I.and Gelade, G. (1980) A feature-intgration theory of ae!eneion. Cog. Psycho!.. 12, 97-136. Wundt, W. (1912) An infr0ducfion fo psychology. Lotidon, George Allen. WurtZ, R. and Mohlcr, C. W. (1976) Organization of monkey superior collicuIus: Enhanced visual response of superficial layer ceils. J. Neurophys., 39, 745-765.
1. Jonides, J. and Somers, P. Voluntary control of the allocation of attention in the visual field. Paper
presented at the meeting of the Midwestern Psychological Association, May, 1977. 2. Jonides. 1. On the automaticity of perceptual leaming. Paper presented at the Psychonomic Society meeting, 1979. 3. Benjamin, M. and Jonides, 1. Cognitive load and maintenance rehearsal. Papel presented at the Midwestern Psychological Association, 1981.
151
Cognition, 10 (1981) 151-158 @Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
Getting developmental differencesor studying child development? ANNETTE
KARMBLOFF-SMITH*
.S.mex University (and Max-Planck-lnstitu t fiir Psychotinguistik
The enticing, yet awful fact aboat child development: is that children develop ! Awful, because it has provoked a plethora of studies, totally unmotivated theoretically, accepted for publication in certain types of journal because the results are ‘significant’ -significant statistically since it is indeed easy to obtain differential effects between, say, five and seven year olds, but questionable as to their significance scientifically. Some of these obviously talented investigators could, in my view, gain so much if they were to spend time observing children prior to getting bogged down in the safe quagmire of variations on a theme -the thenie being the task parameters, not the child. My p;)ychology training was in Geneva and, whatever shortcomings there may be in Piaget’s overall theory and formalizations of child development, there are--inter alla-four interdependent aspects elf his undeniable scientific contribution which, directly or indirectly, infiltrate the won-kof his exstudents. First, Piaget’s genetic epistemology kept in the foreground the relevance of child development for the broader understanding of the g.
*Whilst time limits did not make it possible for me to clculatle this short note for comments, 1 should like to acknowledge recent discussions on tile ylse of the computational metaphor in develop mental psychology with Julie Rutkowska and Stephanie Thornton of tthe Developmental Psychology Group, Cognitive Studies Programme at the University of Sussex. Thanks are also due to a team of youngre%urchc:s workingwithB&bel Inhelder at Geneva University, WHIO stimulated my earlier thinkiug about chzldren’s probliem solting. Requests for reprints should be addressed to Annette Karmiloff-, Smith, 23, cbemin de Chantefleur, 1234 Vessy, Geneva, Switzerknd. ‘The term ‘Piaget-ba&ing’ is borrowed from Freeman (198a) and captures beautifully the ascien. tific attitude of most of those indulging in such work.
152
A. Kanniloff-Smith
treme, where Piagetian disciples reject outright any new results or formal discussions which challenge the theory. Piaget’s theory is so broad and. expressed at such an abstract level cf generality that there is almost always a way of shaping the theory to fit new ideas and new data, or vice versa. But that is surely not the point. There are productive ways of disagreeing with aspects of Piagei’s theory and yet making use of others (e.g., Cillieron, 1977; Karmiloff-Smith, 1979~2:1 l-l 9) or by creating rich, personal interpretations of its meaning (e.g.,.Papert, 1980). Whether this should be considered as a challenge to the theory or as having hitherto gone unnoticed in some obscure footnote of Piaget’s writings .LSfrankly irrelevant in the prese.nt state of the art since, like it or not--depending on whether you are a disciple, a basher or neither -there is no alternative theory of su:,h scope on today’s developmental mark& Wow do I look upon my own work ~nb Lfrzutility, as the Editor of this Journal has requeged ? Five general cc-&r;l$:lations have guided the work since I began experimenting with children in 1973. First, my accent has not been on whether or at what developmentaf stages children do or do not acquire a specific conceptual or linguistic structure. Rather, just those aspects of behavior have been selected for study which are already in the child’s repertoire, with the hope of understanling their changing functions, i.e., their psychological status for the child. This has led to using tasks in which children are successful across various a,<e groups, and thus to attempting to unearth different underlying processes to explain seemingly identical external behavior. My emphasis has been procesr oriented, raising questions about the lzow and the why of the mechanisms inv~>lvedin children’s ongoing behavior in real time. Second, due certainly to m;;/ Genevan backgrounz 1 ! have made extensive us,e of observational data (e.g., my own child re&;liely trying to balance a steel knife on the edge of hi?r plate at mealtime, which led to the hypotheses, experimental design a& analyses reported in KarmiloffSmith and Inhelder, 1974/5; the self repairs in the use of pronouns in afterschool narratives, which led to Ihe hypotheses, experiments and model proposed in Karmiloff-Smith, 1981a), as well as of exploratory techniques (e.g., induced introspection generating epilinguistic comments by children on their spontaneous linguistic self-repairs during experimentation, which led for instance to th.e hypotheses developed in Karmiloff-Smith, 1979~ and 1981~). Both observational data and exploratory techniques are excellent material for formulating hypotheses but, as an experimental psychologist, I have also endeavored to create situations which can actually render the hypotheses and the counter-hypotheses empirically testable (e.g., Karmiloff-Smith, 1980). But the next step, in my view, should always be a return to the observational and exploratory methods of data collection, rather than directly to new ex-
Gettingdevelopmentaldifferences or Jtudyingchilddevelapmen*?
I53
periments. Even when quantifiable datl; have been collected, it was frequently the additional exploratory techniques which shed greater light on underlying processes (e.g., Karmiloff-Smith, 1979~2). Shocking as it may appear to some develogmentalists, I have not been overly worried about criterion-test statistical design, because I hold the conviction that all behavioral manifestations are potentially meaningful to the observer and often should be exploited. Quantitative analyses, even if highly significant statistically, can miss out on important clues to processes. For example, if the first two or three of all sub+s’ responses to, say, fifty items are different from the 48 other respar -es, this is lost in the quantitative analysis,, and yet we may have there just the clue to subjects’ normcll processing procedures prior to their discovery of the expei?menter’,r problem and their use then of experiment-generated behavior (e.g., Karmiloff-Smith, 197&r, where, quantitatively speaking, oldest children’s procedures for French gender attribution look as if they are syntactic in nature but the children’s initial, spontaneous heuristic is clearly a morphophonological procedure). Third, I have not tried to ‘clean up’ or ‘narrow down’ the tasks to get at some so-called pure aspect of the behavior under study. In doing so, it is my view that investigators have again freiluently sidestepped just those revealing indices (the very experimental parasites of which they have rid themselves) about underlying processes and function For example, if, whilst involved in tasks testing usage of the definite article, children are obliged to sit on their hands to avoid pointing, the investigator misses the fact that the definite article is actually functioning as a deictic for the small child (see critique in Karmiloff-Smith, 1978b, and Karmiloff-Smith, 1978~ on the general issue of the importance of a functional analysis of morphemes). Moreover, in narrowing down experiments without returning frequently to observational data, one also runs the risk of seiecting the wrong unit of analysis, e.g., decontextualized sentences versus spans of connected discourse, for the particular aspect of behavior being studied (see KarmiloffSmith, 1977, 137&z and 1979a on the different procedures children use for interpreting articles in sentences and in discourse, and Karmiloff-Smith, 1980 and 1981~~regarding the dynamic role of the use and non-use of pronouns in the establishment of intra-discursive thematic relations; see also the difference between Karmiloff-Smith and Inhelder, 1974/s and KarmiloffSmith, 1981b with respect to different levels of accessing knowledge, e.g., block-balancing as a direct goal versus the same behavior embedded as a means for reaching another goal). In avoiding the ‘narrowing down’ process, I have developed rich task environments, but which necessarily involve ‘noisy’ data. This has led, fourth, to placing major analytical emphasis on what many other researchers might conrider to be marginal data, e.g., microdevelopmental changes during a session, spontaneous repairs, hesitations, distrac-
154
A. Karmil~fff-sm~
tions, etc., often involving indepth analyses of individual protocols (e.g., Karmiloff-Smith, 1979d on ths identification of the fqllowing processes underlying children3 behavior in building closed railway circuits: end-statedriven, procedure-driven, current-state-driven, heuristic-seeking and theorybuilding processes). Fifth, as can be seen from the above, my studies have be. concurrently concerned with both language acquisition and with the processes underlying children’s behavior in non-linguistic problem-solving tasks, which has tended t be a mutually constraining (and exhausting!) exercise, both theoretically dnb experimentally (e.g., Karmiloff-Smith, 1979~ where analogies are suggested between language acquisition and children’s creation of external memory devices). Whilst each of the above points taken separately may read like truisms to some researchers, 1[believe that it is in the combination of the five that resides the utility (and, to be sure, the shortcomings !) of my work in developmental psychology. What are the principle hypotheses which I have entertained with respect to child development? First, a central argument has been that children go beyond successful linguistic expression of their semaniic representations, beyond successful communication with addressees, and beyond successful goal achievement in problem solving. Thus, whilst I endorse the accent which has been placed on the constraints of the development of non-linguistic cognition, as well as on sociodialogic interaction and feedback from failures (e.g., mismatch to adult model, communicative pre;;sures, goal attainment blocked, etc.), these can in my view only be partial and rather general explanations of the mechanisms underlying developmental change. Modifications in children’s behavior, linguistic or non-linguist& cannot be explained by external influences alone. A wide variety of changes, both microdevelopmental and macrodevelopmental, can only be given adequate explanation iif one invokes the hypothesis that children work on their internal representations fir themselves.My argument is that children initially work 3n the mapping between a particular linguistic output and the particblar context to which it most suitably applies. Similarly, they work at a particular, isolated problem-solving procedure and the particular goal sought. However, in each linguistic category in which this mapping becomes automatized, and each time a well functioning problem-solving procedure becomes automatized, then children work on the systems ttsmselves (the intralinguistic system in the case of language, the interrelations between procedures in the case of problem solving). Thus, what .was functioning for the child as a tool progressively becomes part of the problem space itself. Language acquisition can thus be subs’iEz;ed under the general issue of human problem solving. I have argued that very early in development, children take into consideration the formal, and not just the semantic, aspects of
Gettingdevelogmentaldifferences or studyingchilddevelopment?
155
language. For example, the early addition of morphological markers such as nominal and verbal determiners (e.g., articles, the progressive r,-esent tense in English) are added by children at a time when these morphemes contribute nothing to the communicative setting, since interaction then is still deeply context rooted. Thenew markers cannot be explained either by the child’s efforts to match adult models because this merely begs the question of why a particular (often unstressed) marker is added by the child versus so many other potential ones. I have thus submitted that nominal and verbal determiners are initially added by the child for herself (Karmiloff-Smith, 1977, 197Qa). This re&cts the child’s formal rather than merely semantic organizational activities on the consistent aspects of the input and thus allows the child to have some predictive power over the flow of discourse. Another illustration is seen in French children’s attribution of gender to nouns which is initially jkwmal in nature, i.e., via morphophonological oppositions on word endings, and not with respect to the semantic attri&u.tes of referents (Karmiloff-Smith, 1978a). However, it should not be presumed from early correct usage that the articles, say, retain the same function over development. On the contrary, tasksset up specifically to analyse the functional use of articles and other determiners showed that the latter not only change functions over development, but that they gradually acquire plurifunciional status for the child (Karmiloff-Smith, 1979a). However, it must be stressed that the input model itself cannot inform the child of the plurifunc;ional status of its elements used in different contexts. There are no systems ‘in’ the input the child hears, just as there are no explanatory systems observable ‘in’ the physical world. It is only by the child’s own work on the linguistic and other inputs as formal problem spaces that she can create systems and draw benefit from their plurifu .lctional status in encoding and decoding. In the majority of my studies cited above, emphasis has been placed on examples of behavior where children begin a task adequately from the observer’s point of view but subsequently, either during the session or across development, children introduce revealing modifications. There is a cle,ar tendency for children to move from using forms which potentially carry all the necessary information, to spelling out externally, by unifunctional markers, all the pieces of information which were implicit to the observer in their earlier use of an adult-like form. Rather like Bowerman’s late-occurring errors (Bowerman, 1979), the changes are subtle indications to the observer of the child’s reorganization of internal processes. This argument was initially made in Karmiloff-Smith (I 977), at some length in a comparison between language comprehension at the morphemic and the discursive levels (Karmiloff-Smith, 1979b), as well as in a comparison between language llroduction and children’s spontaneous r=iz,:ion of external memory devices in a ma.p
reading task (Karmiloff-Smith, 1979~). In each case, children’s modifications were spontaneous, despite the fact that their earlier behavior was already adequate for successful communication, correct model matching or successful achievement of a goal. Children thus appear to be striving for the delicate balance between encoding and decoding efforts which will ultimately lead to felicitous representational systems. ht first sight, the above looks like the well-known W-shaped phenomenon. But the very concept of a U-shape is from the observer’s viewpoint. It frequently misses the fact that the processes underlying the seemingly identical belravior are quite different; If the underlying processes were the same and if the child had already made the cross-procedural analogies in the first phase, there seems to be no plausible explanation for the changes introduced sponraneously in the intermediate phase. The only feasible reason for these changes in my view is that internal reorganization is taking place and that the identical third phase behavior actually stems from powerful procedures intorlinked into a systemic network. Thus, first children build up isolated, but efficiently functioning procedures. It is possible to keep them isolated, ‘uncontaminated’, because the child as yet entertains no ‘theory’ abour how they function nor does she attempt linkages across procedures. Keeping procedures isolated and repeating successful ma ping operations enables the child to consolidate and render automatized each separate one. Then the child is ready to step up to what I have called a metaprocedural level. In order to have a handle on the procedure the child ‘m6loses her theories on the input, overgeneralizes (and thereby both simplifies and unifies incoming data by ignoring complicating factors), in some cases actually creates ‘observables’ where they should exist according to the currently held theory (e.g., Piaget, Karmiloff-Smith and Bronckart, 1978), spells out explicitly what was implicit in earlier behavior. However, whilst simplified procedures allow children to get a grip on reality (the input or their own output), there is always a trade-off between the simplification and what it then entails (e.g., in discourse production, a simplified process of preempting the initial utterance slot for the thematic subject of 2 narrative forces the: child into tricky lexical choices and self repairs-see Karmiloff-Smith, 1980 and 198la). It is less the search for greater economy which pushes children forward, as we have seen from the examples cited, but rather the search for greater control and mobility between information content and information processing effort, whatever the sphere. In my view, the 1980’s will gain from deeper comparisons of both microand macro-developmental changes in the processes underlying children’s behavior. Comparisons between children’s use of different types of semiotic codes may be instructive in understanding the more generaZ.aspects of repre-
Getting
developmental differences or studying child development?
157
sentation and the specific constraints of particular systems, such as language By contrast, I am convinced that the 1980’s will have nothing to gain whatsoever from further ‘Piaget-bashing’ or further ‘Piaget-confirming’. Rather, there is a dire need in developmental psychology for more profound and precise theorizing. However, to go beyond the purely speculative, it is essential to devise techniques which ‘eliternalize’ subtle aspects of children’s behavior so as to keep the distance fairly close between the theoretical postulates and the ‘observables’ of actual developmental phenomena.
References Bowerman, M. (1979) Starting to make mistakes: the onset of ‘late’ speech errors in language learning children as evidence for ongoing semantic systematization. Ts! A:+ Universi’:y/MIT Workshop. Workshop on U-Shaped Behavioral Growth, June 17-24. Freeman, N. (1980) Strategies of representation in young children: analysis of spatial skills and drawing processes. London, Academic Press. GilliBron, C. (1977) Serial order and vicariant order: the limits of isomorphism, Archives de Psychologie, 45,175,
183-204.
Kaxmiloff-Smith, A. (1977) More about the same: children’s understanding of postarticles, J. Child Lang., 4, 377-394.
Karmiloff-Smith, A. (1978~) The interplay between syntax, semantics and phonology in language acquisition processes, in R. N. Campbell and T. Smith (eds.), Advances in the psychology of language-language development and motherqhild interaction. London, Plenum Press, l.-23. Karmiloff-Smith, A. (1978b) So,.le aspects of the child’s construction of a system of plurifunctional markers. Paper given at Salzburger Beitrage fiIr Lirtguistik V. Karmiloff-Smith, A. (1978~) Adult simultaneous interpretation: a functional analysis of linguistic categories and a comparison with child development. In D. Gerver and W. Sinaiko (eds.), Lunguage, interpretation and communication. London, Plenum Press, 369-384. Karmiloff-Smith, A. (1979a) A fitnctional approach to child language, Cambridge, U.K., Cambridge University Press. Karmilofff-Smith, A. (1979b) Language as a formal problem space for children, unpublished paper presented at the MPG/NIAS Conference on Beyond description in child language research, Nijmegen, June 1979. Kumiloff-Smith, A. (1979~) Micro- ond macrodevelopmental changes in language a#sition and other representatinnal systems, Cog: Sci., 3, 91-l 18. Karmiloff-Smith, A. (19794) Problem-solving procedures in children’s construction and representations of closed railway circuits. Archives de Ipsychologie, KIVII, 180, 33-59. Karmiloff-Smith, A. (1980) Psychological processes underlying pronominahsation and non-pronominalisation in children’s connected discourse. in J. Kreiman and E. Ojedo (eds.), Papers from the pamsession on pronouns and anaphora, Chicago, Chicago Linguistics Societ;;, April 1980,231250. Karmiloff-Smith, A. (1981a) The grammaticai marking of thematic structure m the de;relopment of l&age production. In W. Deutsch (ed.), 7%e child’s construction of lang%xe, Lomdo& Academic Press.
158
A. Ktumiloff-smith
KarmiIoff-Smith, A. (19816) Modifications in children’s representational systems and levels of .accessing knowledge. In B. de fielder (ed.), Knowledge and representation, London, Routledpe and Kegan Pati, Z;armiIoff-Smith, A. (198lc) Using metahnguistic data for a process-oricated approach to language acquisition. Manuscript submitted for pubIication, Max-PIanck-Institut fur Psycholinguistik. Karmiloff-Smith, A. and Inhelder, B. (1974/S) If you want to get ahead, get a theory, Cog., 3, 195212. Papert, S. (1980) Mindstorms:childFen, computers. and powerjkl ideas. England, U.K., The Harvester Press. Piaget, J. (1936) Ltr naissunce de Z’inttWigence chewZ’enfant.Neuchitel, Delachaux et Niestle. Piaget, J. avec KarmiIcff-Smith, A. et Bronckart, J. P., (1978) f_X&alisatiors relatives h Ia pression et i Ia r&&urn. In J. Piaget (ed..), Recherches SWla g&t%alisation.Paris, Presses Universitaires de France, 169-191.
CognitiIon, 10 (1981) UP--i66 @ Elsevier Sequoia S. A., Lausanne - Printed in The Netherlands
1!59
Children’s thinking: What:never develops? FRANK KEIL” Cornell University
Typically, investigators of cognitive development have adopted a perspective opposite to my title (e.g., Siegrer, 1978). Differences between children and adults are explored and descriptions are put forth of how knowledge structures and processes of one age group differ from those of another. There has been surprisingly little attention to those properties that remain invariant across developmental change or, put differently, to the boundary conditions that limit the nature of the change. This perspective may help to explain why the most popular theories of cognitive development-those of Piaget, Bruner, and Vygotsky-propose radical restructurings as children progess through various stages of development. Given such accounts, it often seems as if there is no continuity to develop mental change. One feels a bit like the motorist who asks a rural farmer for directions to a nearby town. The farmer starts giving one set of diu:ctions, pauses, corrects himself and gives another sei, then pauses again for a loriz~r period and finally says, “Nope, can‘t get there from here”, Many of our de-velopmental theories leave us s$randed. in a similar manner. We may have detailed descriptions of children% knowledge structures at several ages but may have no idea of how to get from one stage to another. While some (e.g., Fodor, 1972, 1975; Mandler, 1981) have sharply, ‘criticized theories of this sort on the grounds that the lack of continuity is counterintuitive, nonetheless, these theories represent the predominant point of view ,in cognitive development research. Part of the reason for this theoretical orientation is that frequentiy the only alternative is seen to be traditional learning theory views. Such views do provide continuity by proposing that knowledge is acquired through basic, universal principles of learning that apply in the same manner to all types of knowledge in all learning situations. They do so, however, at the cost of being hopelessly inadequate to explain the acquisition of certain special types of knowledge, such as that of a natural language. *I am gratefulto Jim Cutting, Carol Krumhansl, and K&&i Lockhart-Keil for comments on e,ulier drafts of this paper. Preparation for this paper and some of the research described therein were sup ported by NSF grant BNS-78-06200 to the author. Reprint requests should be sent to Dr. Frank Keil, Psychology Department, Uris Hall, Cornell University, Ithaca, New York 14853. U. S. A.
160
F. Keil
‘I’herc is, however, a thud view of conceptual change that arises out of Chomsky’s work on syntax (e.g., 1965, 1975) but which has been rarely extended to other aspects. of cognition. Chomsky argues that knowled.ge is subject to certain constraints that make some concepts more natural than others and thus easier to acqutie (cf., Osherson, 1978). The notion that there must be a priori constraints for knowledge acquisition to proceed successfully is not a new one, but Chomsky’s work is the first to provide detailed examples of what such constraints are likely to look like. Since even traditional learning theories also assume at least some highly general constraints, a second, distinguishing assumption of the Chomskyan approach is that these constraints are quite restrictive and that they are domain specific. This means that the formal properties of naturai concepts in different cognitive domains should differ in interesting, principled ways and that they should exclude a relatively large number of possible knowledge structures in each dor.uun. I have argued elsewhere (Keil, 1981) that there is increasing evidence fcr such domain specific constraints in several different areas of cognitive development. Wexler and Culicover’s work on syntax (1980), Gelman’s work on number (Gelman and Gallistel, 1978),Osherson’s work on deductive reasoning (1977), and my own work on ontological knowledge (Keil, A979) all suggest that young children, while in many ways dramatically different from adults, nonetheless honor the same highly restrictive constraints throughout their development. Still other examples would include the child’s knowledge of causal relations {Bullock, Gelman and Baillargeon, in press) and organizing principles of infant perception (Bower, 1979; Spelke, in pressj. A constraints-oriented approach provides contiiuity to developmental change in each domain and makes otherwise elusive issues such as whether a particular change is qualitative or quantitative more meaningful. When investigators attenpt to describe conceptual change in terms of unconstrained theories that’ are so powerful that they can describ:? virtually any natural or non-natural knowledge structure (scme,suchas Dresher and Homstein [ 19761 would put scripts in this category), it becomes difficult to know whether children’s structures differ from adults’ in any principled way or not. Tilese considerations point towards a need to uncover and state explicitly the fo.rmal ~nstraints on knowledge for different cognitive domains. Without such speci%ations, one has an incomplete account of cognitive development. But the future specification of domain specific constraints will only partially satisfy the nzxds of cognitive theories, for there are also more domain general constraints and some aspects of knowledge that ‘are not nearly as tightly constrained as others. One of the fundamental problems for future; research is to gain a better understanding of the different roles that weakly and strongIy
Chlidrt&“Jthinking: Whatnw~r develops?
161
constrained knowledge play in cognitive development and how they r Aate to the issue of domain specificity. I have become convinced of the need for such an account through my own work on semantic and conceptual development. Initially, I focused on an aspect of knowledge that is tightly constrained: ontological knowledge and its natural langus~e counterpart predicabiiity, or knowledge of what predicates can be sensibly (but not necessarily truthfully) combined with what terms. Predicability relations can be represented as hierarchical tree structures that obey the M constraint, which simply states that the structtres must be true hierarchies with no downwardly convergent nodes. Predicates dominate terms in this structure if they can be sensibly applied to them, with the topmost node containing predicates that can apply to anything (e.g< , “is interesting”, “is thought about”). Lower nodes contain predicates th;l apply to such categories as physical objects (e.g., “green”, “heavy”) or animals (e.g., “asleep”, ‘excited”) or events (e.g., “was an hour long”, ‘“happened yesterday”). The entire adult tree has approximately fifteen such nodes, each corresponding to a different ontological category. The M constraint is not domain general in that knowiedge structures based on truth vahle (a predicate: dominates a term iff it’ is true of it), instead of predicability, usually do not honor it. There are several ways of generating predicability trees for subjects, the most common of which involvc;s the use of anomaly judgments. Adults all have very similar structures. More interestingly, studies with more than 300 children and in two different languages demonstrate that, even though the children’s trees might be dramatically different from adults (e.g., having only three nodes instead of fifteen), the trees all develop according to precise patterns and obey the M constraint from the earliest stages of tree differentiation. These studies suggest that children’s semantic and conceptual development is rigidly constrained throughout its development by hierarchical predicability relations. A problem arises, however, when one considelrc: the broader issue of how children acquire word meanings. Ontological knowledge suggests that concepts are organized in a rigid dichotomous manner, where certain criterial attributes (or necessary entailments) are either present or not. Rosch’s work (e.g., Rosch, 1978), however, suggests exactly the opposite view of conceptual structure, where few attributes are criteria1 and where concepts are organized around prototypes. Although the limitations of prototypes are being increasingly discussed (Osherson and Smith, 198 1; Armstrong, Gleitman, and Gleitman, Reference note 4), there is no denying that they must play some role in the structure of natural concepts. The *apparent contradiction between these two views of sem,antic and conceptual structure c&n be resolved if they are viewed as working in a compli-
162
F. Keil
mentary manner in the acquisition of word meaning. There appear to be two different levels of word meaning: a defining/characteristic level and an ontological level. There are a number of reasons why defining and characteristic features are at a common level different from the ontological level (Keil, Reference note 2). These include differing behaviors under linguistic hedges, and the differences between contradictions and anomalies. Onotological knowledge has strong domain specific constraints and is not easily accessible to introspection. Defining/characteristic knowledge is more accessible, has more domain general properties, and is not so uniquely constrained. To illustrate the two levels, consider the lay person’s definition of a word such as “mailman”. The person will probably give a defining feature, such as “delivers maE “q and a number of characteristic features, such as typical appearance, manner of delivery, and the like. The lay person will never go on to comment that mailmen have mass, are capable of growth at some points in their life cycles, and breathe. But this lack of mention does not mean that such ontological properties are not essential to that word’s mtaning, as can be seen from anomalies that pose situations inconsistent with such properties (“The mailman was transmitted over the phone line”). A series of studies, some of which are still in progn:ss (Keil, Reference note 2), suggests how the two types of knowledge are involved in the acquisition of word meaning. Ever the youngest of children. appear to have at least some degree of strictly constrained ontological knowledge, even through it may be much less differentiated than the adult form. This knowledge enables children to make a number of inferences about the meaning of a novel word after only hearing one or two predications of it. Moreover, children‘s trees differentiate in such a manner that the children’s inferences from impoverished trees are rarely incorrect but are Imerely too vague in that they don? infer as much as is possible (Keil, Reference note 3). It is quite clear that young children do use this knoilvledgc to make inferences about the meanings of unfamiliar words. If a child hears “The blipple is asleep”, the child will immediately make assumptions about the onl:ological status of blipples and attribute to them certain ontological properties (i.e.., those of the category of animals), Because ontological knowledge is organized as a strict hierarchy, it is an optimally effickt structure for guiding inferences (Simon, 1969). Ontological categories are broad in scope and clearly many further refinements of meaning must occur within their boundaries. This is where characteristic and defining features are essential. Young children are extremely sensitive to familiar exarnplars in their concepts of word meaning and tend to use words as if they are almost exclusively composed of characteristic features. This sort of pattern has been observed for many years. Vygotsky (1965), for
children ‘sthinking: Whatmver develops?
153
example, discussed how the meanings of kinship terms such as “aunt” are originally construed in terms of familiar exemplars (e.g., an aunt is someone about the age of one’s mother who brings presents and visits on weekends). Only later are more defining features (e.g., an aunt is one’s parent’s sister) used. I have been exploring the characteristic-to-defining shift and how it relates to the development of ontological knowledge (Keil and Carroll, 1980; Keil, Reference note 2). One important finding is that even though a child may have a highly instance bound version of a concept that deviates radically from adult usage, the underlying ontological assumptions are usually very sound. (A child may refuse to’believe that an old lady is an aunt but will always know thiit an aunt is a person). In sum, there are two different developmental patterns. There is a tightly constrained differentiat+n of ontological knowledge that, throughout its development, affords reliable inferences about word meaning based on simple predications. Simultaneously, within each of the ontological categories thus indicated, there is a gradual shift away from knowledge solely specified in terms of familiar exemplars to forms requiring more principled, analytic definitions. This second pattern is n-ore weakly constrained by domain general principles, such as prototype structures and semantic congruity effects (prototypes and semantic congruity eft+s are observed in several other domains, such as visual pattern perception &Posnerand Keele, 1968; Banks, 1977 ] ). It also reflects more directly the vagaries of each child’s individual experiences. I do not mean to imply that these are the only domain specific and domain general constraints on semantic and conceptual structure. Clearly there are several others, such as those on aspect (Dowt;r, 1979). It seems likely that these two aspects of word meaning have analogues in other areas of cognitive development. Knowledge acquisition in any domain may frequently employ two types of systems. One system consists of tightly constrained knowledge with its own unique constraints that enable induction to proceed in an efficient fashion. This sort of knowledge is usually not easily accessible, is relatively unchanged by normal environmental variations, and involves intuitive, coherent domains of cognition (such as language, number, music, etc.). It provides a framework within which much finer, more environmentally sensitive differentiations occur and where the principles of organization are less constraining and are domain general. A consequence of this domain generality is that the principles also hold for many completely artificial types of knowledge. This last point about artificiality may explain why cognitive psychology has tended to ignore domain specific constraints. If the phenomena being
164
F. Keil
studied are the learning of arbitrary sets of rules or concepts composed of arbitrary ldisjuncts of meaningless features, it is not surprising that the only constraints uncovered should be domain general. Why should humans have evolved specific constraints unique to a ty i;s of knowledge that has little meaning or use for them? It makes sense that such knowledge should be acquired’ through very general principles of learning and represented in very general ways, since that would be the only mechanism of learning available. Historically, animal learningresearchused to have such an emphasis, where for years investigators examined domain general and weakly constraining principles of learning in artificial tasks involving the pairing of arbitrary stimuli configurations with arbitrary t lponses. Recently, however, the field has undergone a fundamental shift in perspective as increasing attention has been paid to such phenomena as associative bias and preparedness (Seligman, 1970). It was discovered that animals are in fact constrained in all sorts of domain specific ways in terms of how they acquire new information, but such constraints only become apparent when the domain of inquiry is more of a natural unit for the animal (e.g., the learning of taste aversions). Future research on cognitive development must recognize the importance of both types of constraints and, consequently, bol ypes of knowledge. There has existed a bias to ignore domain specific constraints even though these provide the fundamental frameworks that make initial knowledge acquisition possible. Consequently, there h;ave been few cases where we have detailed accounts of how the two types of knowledge might be involved in the acquisition of kltowledge in a given donLain. It is interesting, however, that in the one case where grecise domain specific constraints have be.en formulated, there are suggestions that other aspects of syntax may be learned through quite general feature correlation detectors that generate prototypeiike structures (e.g., Maratsos and Chalkley, 1980), thus illustrating the need for both types of systems. Part of this research enterprise will re’quire a general account of what sorts of knowledge are tightly constrained and why. For example, most knowledge that is usually acquired through explicit instructions (e.g., knowledge of board games) is unlikely to have domain specific constraints, while knowledge that is acquired more naturalistically (e.g., language, moral reasoning, deSthetiCS) does honor such constraints. Obviously, so-called unconstrained knowledge, such as that of board games, must be acquired through the use of systems that are tightly constrained, such as language and spatial representational systems, and will thea*efore be constrained itself. But the important point is that there will be no unique constraints for the domain of board games that distinguish them from, say, the rules of traffic.
Clclildren ‘s thinking: Whatnever develops?
165
Clearl:y these issues are not restricted to the acquisition of knowledge. Fully mature systems must also have both strong domain specific and weaker domain general constraints (there are reasons why strong domain general constraints are unlikely [ Keil, 198 1I ). In adults, however, the systems are often so complex and interactive that their structure may be very difficult to determine. This is where conceptual change, be it developmental, cross-cultural, or historical, provides a valuable tool of inquiry. ?vloreover, in the case of development, the constraints that arc uncovered provide a framework within which to understand how induction could ever succeed and how there is continuity throughout otherwise dramatic change. It may seem paradoxical that we must discover restrictions on what children can learn to be able to explain how they do learn, but, in fact, that is precisely the case. References Banrs, W. P. (1977) Encoding and processing of symbolic information in comparative judgments. In G. H. Bower (ed.),Psychology of Iearningand motivation, Vol. 11. New York, Academic Press.’ Bower, T. G. R. (1979) Human development. San Fransisco, W. H. Freeman. Bullock, M., Celman, R., and Baillargeon, R. (In press) The young child’s understanding of causrl relations. In W. J. Freeman (ed.), The developmental psychotogv of time. New York, Academic Press. Chomsky, N. (1965) Aspects of the theory of syntax. Cambridge, MIT Press. Chomsky, N. (1975) Reflections on language. New York, Pantheon. Dowty, D. R. (1979) Word meaning and Montague grammar. Boston, D. Reidel. Dresher, B. E., and Hornstein, N. (1976) On sone supposed contributions of artificial intelligence to the specific study of language. Cog. 4, 321-398. Fodor, J. A. (1972) Some reflections on L. S. Vygotsky’s Thought and ianguage. Cog. I, 83 -95. Fodor, J. A. (1975) The kmguage of thought. New York, Thomas Y. Crowell. Celman, R., and Gallistel, C. R. (1978) The child’s understanding of number. Cambridge, H‘arvard University Press. Keil, F. C. cL979) Semantic and conceptual development: An ontological perspective. Cambridge, Harvard Unversity Press. Keil, F. C. (1981) Cons&tints on knowledge and cognitive development. Psychof. Rev., 88, 197-227. Keil, F. C., and CarroIl, J. J. (1980) The child’sconception of “tall”: Implications for an alternative view of semantic development. Papers and Reports on C7tik.iDevelopment, 19, 21-18. Mandler, J. M. (1981) Structural invariants in development. In L. S. Liben (ed.), Piaget and the jiiundations of knowledge, Hillsdale, NJ, Lawrence Erlbaum. Maratsos, M., and Chalkley, M. A. (1980) The internal language of children’s syntax: The ontogenesis and representation of syntactic categories. In K. Nelson (ed.), Children’s language (Vol. II). New York, Gardner Press. Osherson, D. N. (1977) Natural connectives: A Chomskyan approach. J. math. Psychol., 26, l-29. Osherson, D. N. (1978) Three conditions on conceptual naturalness. Cog. 6, 263-289. Osherson, D. N,, and Smith, E. E. ;1981) On the adequacyofprototype theory as a theory of concepts. Cog., 9,35-58.
i66
IT Keil
Posner, M. I., end Keele, S. W. (1968) On thegenesisofabstractideas.l exper. Psychol., 77,353-363. Rosch, E. (1978) Principles of categorization. In E. Rosch and B. B. Lloyd (ed.), Cognition and categorization. Hillsdale, NJ, Lawrence Erlbaum. Seligman, M. E. P. (1970) On the generality of laws of iearning. Psychol. Rev., 77,406-418. SKgler, R. S. (ed.), (1978) Children’s thinking: What develops? Hillsdale, NJ, Lawrence Erlbaum, Simon, H. A. (1969) T;le science ofthe artificial. Cambridge, MIT Press. Spelke, E. S. (In press) Perceptual knowledge of objects in infancy. In J. Mehlet (ed.}, Perspectives in cognitive psychology. We&r, K., znd Culicover, P. W. (1980) Formal prirwiples of language acquisition. Cambridge, MIT Press. Vygotsky, L. S. (1965) Throught and bguage. Cambridge, MIT Press. Refmnce Notes 1. Armstrong, S., Gleitman, L. R., and Gleitman, H. What most concepts are not. Forthcoming. 2. Keil, F. C. Two levelsof semantic representation: Implications for a thecry of word meaning acquisition. Unpublished manuscript. 3. Keil, F. C. Inferences from ontological knowledge and word meaniiig acquisition. Unpublished manuscript.
C&&ion, 10 (1981) 167-171 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
167
Early ~settlements in New Cognition WILLIAM
KESEN*
Yaie University
...tnere is onry oozeprinciple (of science) that can be defended under all circumsrances and in all stages of human development. It is the principle: anything goes. Paul Feyerabend in
Against Method
Developmental psyshology, iik;e the rest of the discipline, awoke from its long behavioristic sleep about twenty years ago. The alarm clock may have been the conferences sponsored by the Social Science Research Council Committee on Intellective Processes Research, or Flavell’s book on Piaget, or Brown’s renovation of early language studies, or Fantz’s appreciation ofthe complexity of infant perception, or everi Chomsky’s review of VWM Behaviur. Whatever the provoking event, ilie study of cognitive development has become, over the last score of years -and almost ex&ctly a half-century after Baldwin stated its program in Thought and Things--the core of American research with children. But more has chanr:sd over the years than a cyclical tilt of interests. After reminding you of the clearest changes in the domain of developmental psychology, I will turn to the more fundamental changes in the strategiesof knowledge-acquisitionand then close with a risky forecast of the next decade in the life of Cognition and cognitive development. The changing domain of developmental psychology The liaison with Piaget The story of Piaget’s influence on American psycholog;{ 2nd of the sea.-changes in intellectual affection can be told in near-mythic form by glancing at Piaget’s first three visits to (of all places) New Haven, :,onnecticut. He came first for the Ninth International Congress of Psychology in 1929 when he was perhaps the youngest eminence of the grand assembly that included many of the originators of psychological science. But his precocious fame faded; when Piaget *Reprint requests should be sent to Wlliam Kessen, Psychology Department, Yale Universky, New Haven, Conn. 06520, U.S.A.
next visited New Haven in the early Fifties, his name had dropped from the textbooks and only a graduate seminar in clinical pspchclogy could be found as his aL,lience. How different his visit in 1970, as the furst issues of Cognition were being prepared, when Piaget received an honorary degree from Yale and was greeted with the adulation that is usually reserved for opera Fingers. Att the cycle is sweepin;; down once more and the temper of the field in 1981 is that Piagct has had a (usually) beneficent effect on developmental psychology and that is is now time to move on to more systematic (some dare use the \firord ‘scientific’) studies of the child’s mind. It is an arguable case that we are ‘moving on’ past Piaget without having fully understood the promise and problem contained in his presentation of cognitive epigenesid, but there seems little doubt that the next ten years of child psychology will see the further fading of the Genevan presence.
The metastasis of language study The most dramatic single indicator of the decade’s transformation in the shape of developmental psychology is the outpouring of language studies. When Brown published Wrds arti Things in 1958, he cited less than three dozen references that bore even obliquely on children’s language (fewer than he could find for the language of schizophrenics). It you would seek evidence of the domination of developmental psychology by languajge studies in the ensuing years, you have but to read the present celebratory -issue of Cognition. T;re list of contributors wh3 write about language and language development may be longer than Brown’s 1958 list of references!
The transvaluation of psychology But the affair with Piaget and the burst of language studies are mere signals of a more general phenomenon. Put pl.a.i.Ay,psychology-including developmental psychoiogy-has been redefined as the study of cognition. Friendship has become social cognition, affect is seen as a form of problem-solving, newborn perception is subsumed under a set of transforming rules, iand psychoanalysis is reread as a variant of information processing. Cognition, the feeble infant of the late Fifties and early Sixties, has become an apparently insatiable giant. The recent history of psychology’s redefinition cannot be told in a few pages but an initial move can be made to und, rtand what has happened if we recognize that the fundamental changes that have olccurred over the last two decades have not been changes in domain but rather changes in strategy.
Earlysettlements in New Cognition
169
C?mges in the strategies of know ledge-acquisition The guiding epistemological strategy of psychology from Hehnhoh;i through Hull (to call a name that will not be likely to appear elsewhere in this issue of Cognition) has been based on the domination of generalprocess. Fierce academic battles were fought over whose processes were the correct ones but none of the contestants doubted that someday general laws of psychology would be formulated that would be indifferent to species, age, culture, and (most ambitious claim of all) specl.ic psychological content. The apogee of the conviction that general process would triumph (or even had triumphed) in psychology can be discerned in Hull’s statement in 1950 of ‘A primary social science law’. The exaggerations or the Great Learning Theorists are the easiest to cluck over because they are so plainly and arrogantly stated but we must remember that all other ‘schools’ of psychology adhered, even if less noisily, to the basic principle of process domination. Developmental psychologists should remember, too, that PiEget saw human intelligence ultimately as the ability to explain the largest domain of phenomena with the smallest possible number of formal principles. Over the last twenty years, the universa’!ism of process domination in psychology, tied as it was to prejudices about method and sustained as it was by cult formation, collapsed. And not with a bang; the psychology of prc%ess was succeeded by the psychology of contents almost soundlessly. The transformation is too close to us’ and too unstable to describe in sure detail, but several characteristics of the newer mode in the study of cognitive develop ment are worth brief notice. l&e partitioning of developmental psy‘hology The uniqueness of human language required that, when attention was turned
to language studies in the Sixties and Seventies, the shift from defimition by process to definition by content would take place. Language was of interest as a particular kind of human behavior and, if it did not tit with classical processes, so much the worse for classical processes. And so it has happened with a host of other psychological contents. !-nagery, metaphorical reasoning, emotion, story comprehension, parental artachment, for examples, have flourished as intellectual categories and centers of recent empirical research not because they were useful representations of some general process but because they were (on grounds that have never been formally stated) interesting aspects of the human mind. A hun’dred flowers have bloomed.
I? 0
.lWliurn Kessen
l%e shrugging off of theory
The partitioning of the field of developmental psychology and the end of uniformity of principle have led to a curious insouciance about systematic theory. Although each subdomain is evolving or adapting its rules of evidence and proposing conceptual schemes to hold its data together, the glue that binds the doma.ins in some neigbborly relation derives largely from smart software and the ubiquity of the microprocessor (a name filled with1irony). And therein lies another process-erosive force. The plain fact that common computer programs are far more intricate than any psychlological theory that existed before 1960 and the further fact that the implications of complicated programs can be radically modified during a long day at the terminal combine to make fixed and simple theories unlikely. In an analogy to New England’s weather, ii’?esday’s theory doesn’t make you happy, wait ‘til Wednesday’s. Thefomatlion of smalland exclusiveclubs
With no general theory and a partitioning of the contents of developmental psychology, an interesting cultural construction has been made. Young investigators stake out their own intellectual claim in the field of the new cognition-the ‘settlements’ of my title- and become known as the Expert on Imagery, or the Expert on Paternal Attachment, or the Expert on Intramoda1 Perception, and so on through the agenda of exciting new research in cognitive deveiopment. In contrast to the strongly hierarchical arrangements of the days of process psychology where a Chief (Skinner ‘or Hull or Festinger or Anna Freud) topped an orderly status tree, the emer,ging pattern (so obviously exaggerated in my sparse accot’. :t; is to have almost as many Chiefs as there are domains. Pie growth of a psychology of context NOIinly have we seen a renascence of the psychology of content, with a corresponding loosening of the canonical bonds of shared theory, certain method, and largish schools; there is also at work in the study of children’s minds iu new commitment to a recognition that few.cognitive acts are free of their social and cultural context. The recqgnition is revealed in concern about the generalizability of traditional laboratory designs, in systematic Tttempts to devise powerful empirical procedures different from the classical forms (most recently by Cole and by Bronfenbrenner), and in new moves toward applications of developmental psychology in schools, clinics, and hospitals. The ancient line of process psychology has had little room for the ambiguities and
Ear&
settkments
in New
C”gnitbn
171
eccentricities of a contextual psychology; thus, another support has been given to the psychology of ‘anything goes’. The next decade If we look back over the changes that have taken place in developmental psychology over the last ten years, none of us can be confident as prophets and forecasters. But, a guess or two can be made about the next moments in our history. The new era :,f a psychology of content and context, an era which has unfrozen methods and opened captivating new areas of research for child psychologists, has not reached its full flowering. Over the short future, the number of partitiorls of the field will increase and more settlements along the littoral of New Cognition will be made; it is possible to roresee a renewal of interest in the developmental psychology of the arts (;n:speciqlly of music), in children’s problem-solving in natural settings, in the expressive and poetic uses of language, in dreams, and (at last) in a developmental account of emotion. ‘l;t, with al! the challenge and agitating innovation, there ar-e already signs that the psycl--9logy of content will reach its own redirk.+ing juncture. For one, the partitioning of the field can lead, if unchecked, to esoteric fractions less and less available to the general developmentalist. There is the further risk +I*Ttthe subdomains will becomz victims of psychology’s most t;rsistantail:*cIt, premature refinement. Finally, and of implication hardefst to diagnose, continuing partition may lead the new Chiefs to a narrowness and loneliness that will weaken their stamina. What will not change over the next decade, ev ‘n if developmental psycholhcories, is the contextual definition ogy is returned to domination by grand tllti of the psychological enterprise itseif. MacKenzie and Barnes, in Natural Order, have sa:zi the words in piainest guise: Scientific research remains, inevitably, patterned by particular goal-orientations whose relationship ‘o the structure of the scientific community and the widersociety is always an open question.
References Bar-es, B., and Shapin, S. (eds.), (1979) Natwv’ vder: Hf*wical studies of scient&% culture, bndon, Sage. I eyerabend, P. (1975) Against method. London, Verso.
Cognition, 10 (1981) 173-179 @ Eievier Sequoia S. A., Lausanne -. Printed in the ivstherlands
173
Researchon mental imagery: Some gFcsals and directions STEPHEN MICHAEL Brmfeis
KOSSLYN”
Uniwersity
For almost a decade my colleagues and I have been engaged in a concerted effort to understand how information is represented in, and accessed from, visual mental images (see Kosslyn, 1980). As such, we have been concerned primarily with questions pertaining to how images are generated (as when one forms an image of a novel scene), how images are inspected (as when one counts the number of windows in one’s livingroom by “examining” an image of the room), how images are transformed (as when or&eexrlands, shrinks, shifts the apparent location of, rotates, or otherwise con!,&s an image), and, questions pertaining to when images are used spontaneously in the course of retrieving information from memory (as when one tries to repor? the shape of a German Shepherd’s ears). This work has not been without ?s critics (see Kosslyn, Pinker, Smith and Shwartz, 1979), and the general topic remains extremely controversial (e.g., see PyIyshyn, 1981). In this paper I briefly describe the broad goals of our research program and the general direction in which we hope to proceed. This seems worthwhile if only because some of the current controversy may be the result of misconceptions about the long-range goals and purposes of our research. The approach Our general approach is governed by a particular conception of an adequate theory of imagery and how best to collect data that will aid in formulating such a theory. Form
of
the theory
Any theory must specify the nature of thL theoretical entities and the principles of how they interact. The entities in our theory are “functional capacIties” of the brain. That is, we are interested in things the brain can do in the course of representing and processing images. We are not interested iltrcharacterizing the “machine language” of the brain (or the “elementary informa*Reprintrequests shou,ldbe sent to
Stephen M. Kosslyn, Programin Linguistk.and Cognitive Science, Brantbis University, Walthaw Mass., 02254, U.S.A.
174
S. M. Ko&w
tion process&‘) from which all higher-order units (akin to sub-routines in a computer program) are composed. Rather, we wish to under&and the h&herorder units themselves, which we take to be like the modular components of a very large computer program. We have found it useful to divide such f Inctional capaltities into two broad classes, structures and processes; structures serve to reylresent information, and processes manipulate representations (sze chapter 5 of Kosslyn, 1980, for a detailed dnscussion of these notions). We distinguish between two kinds of structure:, media and data structures. Media carry no information in their own rig&, but rather support the informationbearing data structures. Properties of media (such as the “grain” and “extent” of :‘ne “visual buffer” we posit) affect all data structures that occur within them. .A number of distinctions can be drawn between different kinds of processes as well, notably between those that compare two-structures or parts thereof, thas\: that transform a data structure (e.g., by adding or deleting information)., and those that produce new data structures. ‘?ere are two ways of couching the principles of how processes may be conL;atenatll:d, recruiting and producing various struc%.‘es as necessary. First, these principles Fre inherent in the input/output characteristics of each individual proi~;~ssin* component: the ouput from one provides the input to another. a* i a complete understanding of the nature of the individual processes w. c 4 allow one to specify how they would be activated in turn when any giwr~ ask ISperformed. Second, there are principles that can be abstracted away t’::om the interaction of individual processes and structures and stated as generalizations, such as those offered by Marr ;” 976). As one example in the Kosslyn anti Shwartz model (see crlapter 9 of Kosslyn, 1980, for details), “propositional”‘and ‘imagistic” representations are processed in parallel when one: is asked to retrieve facts about objects from memory. It is hgClftant to realize that the actual expression of the theory need not preserve each of the functional capacities as distinct entities. For example, it may turn out that use of Group Theory allows one to express the way image transformations work without reference to the specific transformations (espansion, rotation, etc). Further, there may be different ways of expressing the theory, different “notational variants”, that are more or less useful for specific purposes; using Group Theory may be useful for generating precise predictions; about the rates of transformation, say, whereas some other notation might be more useful for predicting when an incremental transformation (in which images seem to pass through the intermediate points along a tra&ctory) will be used instead of a “blink” transformation (in which images are transformed by ieating an initial image fade and replacing it with a second image of the object which is transformed in some way-see Kosslyn, 1980,
Research on mentalimagery:Some goalsand direct&s
175
1981). In all cases, however, the theory is expressing law-like regularities among a set of functional capacities of the brain.. A research methodology
A major problem in the study of image representation today is to devise ways of systematically collecting data that will place the strongest constraints on our theories. We have found it very useful to consider alternative ways in which one could build a computer simulation model that embodies the individual functional capacities posited by the theory, and then have conducted research to discriminate among these alternatives. TGs approach has proved very tiseful in practice (see Kosslyn, 1980) largely because: 1) it forces one to think very concretely about the formal properties of structures and pra:esses; 2) the functional capacities one considers seem required in order to bu..ld a working model of representation and processing; and 3) each component of the theory is available to bear on any imagery phenomenon, and hence the characterization of each one is to some extent constrained by the way the other functional capacities have been specified. This approach is more than z simple heuristic to collecting data; wc believe that the properties of the theoretical entities we are positing in attempting to provide simple, straighl forward computational accounts of the data do in fact reflect the actual state of affairs in the bram. Extending the them-y Qur current theory of imagery is only a small fragment of a complete theory. In this section I wish to identify the kinds of research that will further the development of the kind of theory we envision. Developing the core theory
Considerable research can be conducted within the context of the current “core theory “. The curre;lt theory has numerous “place holders”, components that have yet to be specified and are only vaguely described. In particular, we need to specify the nature of the long-term memory representat”ons underlying images (are they “depictive”? “propositional”? “object centered” or “viewer centered”?), the principles that determine which parts will be included in an image and which will not (and which determine what image will in fact be formed when one is asked to image some object-such as which dog
176
S. M. Kossiyn
will be imaged when one isasked to image a dog), and the mechanisms underlying various forms of image transformation. Other relevant topics include questions about how motion, color, and 3-D information (see Pinker, 1980) are represented in imagery. In addition, considera’Ay more work is necessary to specify the properties o&and justify positing-the analogue spatial medium in which images purportedly occur. Extending the domain The kind of theory we have been developing can be extended to address more than the original set of questions about how information is represented in, and accessed from, visual mental images. Six directions seem especially promising, as noted below. Ontogenetic development Che measure of the adequacy of a theory of adult cognition is the plausibility of the ontogenesis of the putative structures and processes. That is, the adult state developed over age, and if one must make unreasonable assumptions about the child’s abilities to represent and process information in order to account for development, the theory is’inadequate (cf., Chomsky, 1965). Thus, if only for this reason the study of the child’s abilities to represent and process information in imagery seems a key area for future research, We plan to study the development of the basic components posited by OURrheory and to compare the development of thes,e structures and processes with those underlying linguistic processing. It would be especially interesting if some imagery abilities develop prior to language-#related ones, which may suggest ways of teaching young children more effectively. hdividti
differences
One problem with studying individual differences is that there are too many of &em. One use of a nomothetic theory, hlowever, is to specify the respects in which people can differ in interesting ways. In the kind of theory we envision, the imagedry system will be specified in terms of a set of functional capa&es-an4 people will certainly differ in terms of the relative efficacy of the various c~pa&zsi A characterization of the relative efficacies of the different functional capacities (e.g.; the acuity of the visual buffer, the amount of information that can be maintained at once, the ease of transforming images,
Research on mental im+!gery: &Tome goals and directions
177
etc.) of different people would serve two ends: First, insofar as the abilities could be shown to be relatively independent, this provides further support for the putative distinct functional capacities themselves. Second, a characterizatioiz of a person’s relative strengths and weaknesses would suggest what kinds of strategies that person would find most useful when using imagery in problem-solving and the like. Cheep t represeiatation has long been thought that concepts may be represented in different formats (see Gibson, 1969). There is, however, surprisingly little data on the topic, and virtually none that implicates imagery as a form of concept represc:ntatio:l (see Farah and KossIyn, in press). If images (“surface” or “deep”) are used to represent some kinds of concepts (e.g., perhaps of artistic styles), then many of the constraints imposed by the imagery system will affect how easily such concepts can be learned and used. It
Perception It is clear that imagery and perception share some of the same mechanisms (see Finke, 1980; Shepard and Podgorny, 1978), and it is important a) to specify the nature of the structures and processes that are utilized both in imagery and perception, and b) to examine their respective roles (which need not be the s?me) in the two task-domains. For example, in our theory we posit an active-memory medium (the “visual buffer”) which supports the depictive representations that underlie the experience of “having $0 image” {see Kosslyn, 1980, 198 1). This same medium may be used in perception to support something like the “21/2-D Sketch” posited by Marr and Nishihara (1978), although an image would not be a “2%Sketch” (which consists of extremely local primitives specifying the distance and orientation of each point on a surface relative to a viewer; an image would consist of much larger units, corresponding in some cases to parts of objects or scenes). In this case the structure may be playing similar roles in imagery and perception. In contrast, the role played by the underlying “deep” representations of appearance may be quite different: in imagery these representations are used to generate a depictive representation in the active-memory medium, whereas in perception they may not be used generatively, but rather may be used in object recognition via a matching process. In any event, these are th,: kinds of questions we believe should be explored on this topic.
178
S. M. Kos&n
77zinkingand problem-solving Imagery has often been implicated anecdotally in thinking and problem-solving (e.g., see Shepard, 1975), but has not yet been carefully studied in this role. All visual thinking would seem toinvolve at most four classes of imagery processes, namely image generation, maintenance, transformation, and interpretation. For example, consider a very simple problem, imagining how to rearrange furniture so that all of the pieces would fit against the walls in a room. In this case, 1) an image of the furniture and room must be formed, 2) the image must be held in mind while operated upon, 3) the various pieces of furniture must be shifted about in the image, and 4) one must be able to “see” wheth.er an imaged piece of furniture does in fact fit in a given location. In some cases one need not transfom an image in order to solve a problem, as may occur when images of nested circles are used as an aid to solving classinclusion problems in logic. Note also that in this case the interpretation process now involves more than simply identit’ying the imaged pattern, but also involves recalling the meaning associated with the pattern (i.e., what each circle stands for and what the various topoIogica1 relations among the circles symbolize). Research on imagery in problem solving, then, could focus on discovering how constraints on each of the four kinds of processes described above dictate the most effective way to use imagery to represent a given type of problem. Brain dunage There is intriguing evidence that some of ithe functional capacities we have posited arise through the operation of mechanisms at specific loci in the brain. Bisiach and Luzzatti (1978), for example, report a case in which a right hemisphere parietal lesion resulted in a patient ignoring the left half of his images, just as if the “mind’s eye” only operated over half of the “‘visual buffer” posited in our theory, li: would be relatively easy to devise a series of simple taslq each tapping one of the functional capacities posited by the theory. It would be very strong support for the theory if various functional capacities were independently affected by different kinds of brain damage. And it would be even stronger evidence if we could literally identify different functional capaities with specific neural loci. Findings already in the literature give us caus3 for cautious optimism. Concluding remarks Research on the topics outlined above will serfe two roles: Fiat, in the event that we can successfully extend the theory, this is evidence that the initial
Research on mental imagery: Some goals and directions
179
conceptions were correct. Second, in the course of so extending the theory we will discover new and interesting phenomena that any theory must then address. It is only through this cycle of mking new discoveries in the course of trying to formulate better theories that we eventually will come, I believe, to unravel the mystery of how the brain represents and processes visual infcsmation in memory. References Bisiach, E., and Luzzatti, C. (1978) Unilateral neglect ofrepr$sentational space. Cortex 24: 129-133. Chomsky, N. (1965) Aspects of the Theory of Syntax. Cambnldge, MA, MIT Press. Farah, M., and Kosslyn, S. M. (In prees) Concept deveiopmenl:. In H. W. Reese and L. P. Lipsitt (eds.), Advances ifl Child Development and Behhavior.New York, Academic Press. Finke, R. A. (1980) Levels of equivalmce in imagery and perception. PsychoZ.Rev., 87, 113-13X. Gibson, E. J. (1969)Principles of Perceptual Learning and Development. NewYork, Appleton-CenturyCrofts. Kosslyn, S. M. (1980) Image and Mind. Cambridge, MA, Harvarri University Press. Kosslyn, S. M. f 1981) The medium and th+=message in mental imgery : a theory.Psychol. Rev., 88, 46 66. Kosslyn, S. M., Pinker, S., Smith, G. E., and Shwartz, S. P. (1’179) On the ic;nystification of mental imagery. Behav. Brain Sci., 2, 535 -581. Marr, D. (1976) Early processing of visual, information. Phi!. 7bans. Roy. Sot. V, 275, 483-524. Marr, D., End Nishihara, H. K. (19; 3) Visual information processing: artificial intelligence and the sensorium of sight. Technol. Rev., 81, 2-23. Pinker, S. (1980) Mental imagery and the third dimension. J. exper. Psychd Gen., 109, 354-371. Pylyshyn, Z. W. (1981) The imagery debate: analogue media uerslls tacit knowledge. Psychol. Rev., 88, 16-45. Shepard, R. N. (i 975) Form, formation, and transformation of internal representations. In R. L. Solso (ed.); Informatkm Processing and Cognittin: Thk Loyok Sympuzkx Hillsdale, NJ, Erlbaum Associates. Shepard, R. N., and Poe+*-v, P. (1978) Cognitive processes that .resemble perceptual processes. In W. K. Estes (ea.), hr::.. ..+okof Learning and CognitiveRecesses (vol. 5). Hillsdale, NJ, Erl8aum Associates.
Cognition, 10 (2981) 181-186 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
1811
Logic in infancy JONAS LANGER’ University
of California,
Berkeley
Determining the elementary structures of cognition during infancy and ear1.y childhood is crucial to analyzing the problem of knowledge-its origins, development, and organization (Langer, 1980 ; in press; in preparation). The basic theoretical aim is to formulate a structural development4 theory which is formal and epistemological as well as psychological; and which serves as a heuristic guide to the research as well as explaining the findings. This entails formulating a model of the fundamental generative organization of logical cognition which encompasses both hypothesized elementary forms and hypothesized developmental transformations in these elementary forms. The theory I have therefore proposed is that logical cognition consists of elementary interactive operations (e.g., composing, substituting, and negating), their threefold structural organization (i.e., combinative, relational, and conditional), and their constructed inferential or deductive products (i.e., equivalence, ordlered nonequivalence, and reversibility). The basic research strategy is to investigate, as directly as possible, children’s developing constructive interactions which entail progressive logical operations from ages 6 to 60 months. Accordingly, our research is designed to record, code, and analyze (a) all subjects’ direct interactions with a wide variety of objects and sets, and (b) the constructive transformations which these interactions produce in objects and sets. The objects (e.g., dolls, rings, and blocks) are made out of malleable Play-Doh, out of nonmalleable material (e.g., wood), and out of both malleable and nonmalleable material. Infants’ interactions with and constructive transformations of these objects are analyzed into microunits while at the same time preserving the sequence and context in which they occur. The data include quantitative measures of infants’ spontaneous manipulations, as well a.s detailed qualita-. tive observations. The results of experimenter-provoked manipulations and counterconditions supplement these measures. Two form? of analyses are fundamental to this inquiry. These are analyses of the .part-whole and the means-ends transformations which infants construct in structuring their interactions with their environments.
*Reprint requests should be sent to Jonas Langer, Psychology Department, University of Caiifornia, Berkeley, California 94720, U.S.A.
! 82
Jonas Langer
Part-tihole transformations vary as a function of the objects with which subjects interact. Some are possible with all types of objects. Th.e basic ones are .- (1) uniting objects into sets or series; (2) reuniting sets or series into derivative variants; (3) separating sets or series into either related subsets or subseries; and (4) separating sets or series into unrelated objects. Only the part-whole relations of malleable objects can be transformed from: (5) one form of object into another (e.g., from a ring into a solid); (6) a single object into a set or a series by decomposing the whole into smaller parts; and (7) a set or series into a single object by uniting the elements into a larger whole. Two additional’part-whole transformations are unique to the structure of combined malleable and discrete (nonmalleable) objects. These are (8) attaching malleable and discrete parts into larger conglomerate wholes; and (9j detaching conglomerate whoies into smaller malleable and discrete parts. The development during infancy of these nine basic part-whole transformations is the data base for our inquiry into the origins and early development of logical cognition. As early as age six months infants consistently unite nonmalleable wooden objects into 2-abject compositions, and rarely into 3-object compositions. They all consistently reunite these compositions into derivative recompositions. Some six-month-olds already begin to reunite their 2-object compositions into derivative 2-object recompositions by pragmatic exchange operations of replacement, substitution, and commutativity which produce quantitative equivalence within single sets. Two levels of elementary logical operations develop during infancy. Consider substitution. First-order substitution is limited to single, nonreversible 2-object sets at age 6 months. One third of six-month-old Ss begin to substitute I object for another in 2-abject sets which they have constructed. These infants barely begin to produce quantitative equivalence between consecutive versions of 2-abject sets. They take away 1 element from an initial 2-objec,t set they have just composed and immediately substitute another object in its place. No six-month-old Ss invert the exchange by resubstituting the original object so as to reconstruct their initial 2-abject sets. First-order substitution is expanded during infancy to produce quant?ative equivalence in ever larger numbers of objects within single sets. So, by age ‘18 *months half of the Ss substitute: in as rnany as 4-object sets. Moreover, substitution is becoming progressively reversible by negation; where infants invert their substitution to reconstruct the initial identity set. Yet all first-order operations, including substitution, are limited to producing equivalence, ordered nonequivalence, and reversibility within single sets which 8re not related to any of the infants’ other compositions. The major development during the second year of infancy is the origins of second-order operations. Second-oraer, unlike first-order, operations are
Logic in infancy
183
coordinative part-whole constructions. They coordinate elementary firstorder part-whole structures to each other. This produces a second level of more powerful operational structures such as compositions of compoz,itions and equivalences of equivalences. Consider substitution again. All eighteen-month-old infants compose &wo very small sets in relation to each other. They produce two compositions in simultaneity or in partial temporal overlap. Almost all infants also match two compositions by one-to-one correspondence between the elements such that the two sets are quantitatively equivalent (cf., Sugarman, 1988, in press). Almost half of these infants coherently transform such binary sets by substituting equivalent numbers of objects within the two sets such that they preserve the quantitative equivalence between the two sets while transforming them. Many of these secondarder constructions are already featured by reversibility in which infants negate their initial substitutions by inverse substitutions. It should be stressed that even second-order substitutions still constitute precursory opemtions or proto-operations for many reasons. One illustrative reason is that only proto-operations are always limited to exchanging elements within sets and never fully exchanging elements between sets; fully formed operations are not limited in this way. In no way can our findings be construed as reducing advanced stages of necessary and formal operations, which originate during adolescence, to first- or second-order protooperations, which originate during infancy. Rather, what we are discovering are the structural developmental continuities and discontinuities between proto-operational and operational cognition. Para.llel developments are found in the means-ends transformations generated by infants. Unlike part-whole transformations, meansends transformations are generated by infants when they are constructing or orienting to some goal or objective such as maintaining or reproducing an interesting or desired happning (cf. Piaget, 1952). For instance, the meansends relation between a desired object which is moving away from an inifant (.the end) and an instrumental organ such as the hand rvhich is at rest (the means) is transformed when the organ gropes or searches after the objective. Consider the development of causal meansends functiolns which are central to the formation of physical concepts. First-order causal protofuntitions develop durmg the first two years. They begin as early as; age 6 months with the construction, minimal replication, and observation of effects which are direct functions of causes (e.g., one object is used as a means with which to push another dependent object several times in succession). This is csmplemented by the origins of causal anticipation or prediction (e.g., one object is used to block and stop another object which is rolling in front of the infant).
184
Jmas Longer
There is marked progress in constructing first-ordef or direct causal protofunctions by age 18 months. This includes generating dependent semivariables or effects which are direst functions of independent semivariables or causes (e.g., objects are pushed harder and harder). The direct functional dependency of the dependent semivariables upon the independent semi-.ariables may b,: formalized as one-way ratio relations, such as ‘Moving 1 urther is a function of Pushing Larder’. This is what differentiates firstorder from second-order causal protofunctions which originate during the second year of infancy. Secondorder *protofunctions are coordinative means-ends constructions. They coordinate elementary first-order means-ends structures to each other. This produces a second structural le,vel of more powerful protofunctions. The structural hypothesis is that th; effects are directly dependent upon the causes in f&t-order functions. In ccntrast, the effects begin to be proportional to the causes in second-order functions. That is, the exp) cted structural developmental difference is that firstorder protofunctions are feature< by direct ratio relations while second-order protofunctions are marked by indirect analogical or proportional relations. Eighteen-month-olds, like younger infants, use one object as an instrument with which to push a second dependent object. But beginning at age 18 months, when the effect is that the dependent object 41s away, then infants may also transform the instrument into a means with which to block the dependent qbject. Correlatively, infants thereby transform the end or goal from rolling to stopping. As soon as the dependent object stops rolling, infants transform the same instrumental object back into a means with which to make the dependent object roll away’ again. And so on. Thus, infants begin to covary their transfo,rmations of both means and ends QKindependent and dependent semivariables at age 18 months. These covariations form coordinate preproportional dependencies between causes and effects. These preproportions coordinate previously constructed firstorder dependencies to each other. This produces new second-order dependency protofunctions, such as ‘Moving is a function of Pushing as Stopping-is a function of Blocking’. Or put more simply in its analogical form: ‘Pushing is to Moving as Blocking is to Stopping’. Preproportions, then, are fundamental in permitting the passage from one functional equivalence class to another through a constant transformation (cf., Piaget, Grize, Szeminska, Vinh Bang, 196811977). Here the equivalence classes are transformed by infants from ‘Pushing’ to ‘Blocking’ independent semivariables and from ‘Moving’ to ‘Stopping’ dependent semivariables. The finding that protofunctions and proto-operations develop in parallel from first- to second-order constructions during infancy leads to the
hypothesis that they initiate the structural transformation of prerepresentational into representational cognition. The hypothesis is that the development of second-order constructions is the main source of transforming sensorimotor cognition characteristics of early infancy into semisensorimotor and semi-conceptual cognition characteristics of late infancy and early childhood. The symbolic aspect of developing representation may be marked by extending cognition into the not-here 2nd the not-now. The conceptual aspect of developing representation, I would propose, consists of zransforming firstarder into second-order proto-operations and protofunctions. This proposal provides a formal structural account of the origins of representation during this age period; nlrmely that it is marked by coordinate secondorder constructions. The thesis is that the main d,evelopments in logical and physical cogr;ition up to at least age eighteen manths consist of progress in first-order protooperations and protofulictions and, especially, in their coordination into second-order proto-operations and protofunctions. The main develoyilli;n+s in symbolic and linguistic behavior up to at least this age do not parallel those of cognitive behavior. Typically, infants are still only at the. one-word stage at this age (e.g., Brown, 1973). The working hypothesis is that cognitive developments during this age period are basically independent of linguistic developments. If anything, linguistic developments lag greatly behind cognitive developments up to and im:luding at least age eighteen months; as seems to he the case if we use a?~our standard of comparison current linguistic analyses of semantics and syntax (e.g , Bowerman, 1978). Together these various foci of our research rnake it possible to investigate the developing relations between infants’ logical, physical, and symbolic (including linguistic) constructions with a view to determining the nature of elementary cognitive development. Discovering that logical operations and physical functions originate during the first year of life and progress during the second year to the formation of relatively powerful if still rudimentary second-order constructions opens up many new theoretical and empirical possibilities in analyzing the problem of kncwledgc. To name but one, it brings into serious question long held epistemological and psychological asslimptions about the necessity of symbolization, including language, for the construction of elementary logical operations and elementary physical functions. References Bowerman, M. (1978) Structural relationships in children’s utterances: Syntactic or semantic? In Bloom, L. (ed.),Readings in Language Development. New York, Wiley.
186
Jonas Langer
Brown, R. (1973)A Fkst Lunguage: The Eurly Stuges. Cambridge, Harvard University Press. Ianger, J. (1980) The Srigins ofLogic: Six to Twelve Months. New York, Academic Press. Lamger, J. (In pmparation) The B&ins of Logic: One to Two Yazrs. L-anger, J. (In press) From prerepresentational to representational cognition. In Forman G. (ed.), Action and Thaugk. New York, Academic Press. P&et, J. (1952) K’~EOrigins of Intelligence in Chihfren. New York, International University Press. Piaget, J., Grize, 8.-B., Smminska, A., and Vinh Bang (196811977) Episfemology and Psychology of Functions. Dordrecht, Reidel. Sugarman, S..(1980) Scheme, Order, and Outcome. Unpublished doctoral disseriatlon, University of California, Berkeley. Sugarman, S. (In press) Transitions in early representational intelligence. In Forman, C. (ed .), Action und Thought. New York, Academic Press.
Cognition, 10 (1981) 187-192 0 Elsevier Sequoia S.A., Lausanne - Printed in The Yetherlands
187
WILLEM J. M. LEVELT+ f&x- Planck-Ins titcdt ftir PsychoBnguistik
A decade of Cognition has made it a very established journal. The editor and associate editors presumably have mixed feelings about this: they will not be insensitive to success, I guess, but surely they have a problem with establishment. The original motivation for the journal was, in part, to get away from the masquerade of the standard paper format, which tends to inhibit :he expression of the larger theoretical framework or the scientific philosophy of the author. For znother part, and not unrelated to- this, the journal was in tended to serve as a forum to discuss the utility of cognitive science for society and its (ab)uses in changing society. In the first editorial one can read: ‘Thus it is our duty to discuss not only the practical value of our scientific conceptions in light of the problems faced by people and societies but also to evaluate possibie applications in the light of what we know about ourselves’. Rereading this editorial I couldn’t help being reminded of a similar statement: ‘This peace we will not find before we have changed ourselves; and in order to change ourselves we miJ/ill first have to know ourselves. This knowledge psychology tries to provide us? by patient work; and we will have to wait also patiently for the time: that this knowledge can be utilized in practice’. This was written in tiic year 1909 by G. Heymans, the well-known Dutch psychologist, in a paper Galled ‘The coming century of psychology’, Admittedly, Cognition’s editorials show less patience, but this should be understandable given that three quarters of Heymans’ century have passed by withoui noticeable results for either peace or society. The same fact, however, makes one wary to express opinions about the utility of one’s own work for the coming decade (as I was asked to do). It is not going to be more than that of our scientific predecessors for present SOciety: very limited, if any, Still, the historical perspective is less a source of pessimism for me than one of inspiratinn. Whatever the research 1 will do in the. coming decade, it will stay centered around some fundamental notions and issues which are classical in psychology. My scientific roots are in a continental tradition which is dominantly mentalistic and nativistic. The intellectual climate at Leyden University durirzg my education was eclectic, but with clear overtones of phenomenology, Gestalt psychology and ethology. I *Reprint requests shourd be sent to: W, J. Levelt, MaxPlanck-Institut-Zir Dalseweg 79, Nijmeger:., The Netherlands.
PsycholinguisttL. Berg en
188
Willem J. M. Levelt
vividly remember my surprise at listening to Chomsky’s attacks on psychology during my* -j-w in Cambridge, Mass. 19651966. Psycholog/ apparently was Skinnerian psychology, as it still was in Chomsky’s article ‘Psychology and ideology’ which opens the first issue of Cognition. Honestly, I didn’t even know the difference between classical and instrumental conditioning at the time, in spite of my Ph. D. in psychology. In subsequent years, I have come to rtialize that a major influence on my thinking stemmed from A. Michotte, the Belgian psychologist, who in 1959 had given me a sernester’s hospitality in his laboratory at Louvain university. During this stay he had worked intensively with me, in spite of the 5’7years difference in age. Michotte had been trained by Wundt, and especially by Kiilpe, the fou ader of the so-called ‘Wiirzburgschool’.This school had departed from Wundt by studying ‘higher’ mental processes by means of experimentation. The approach remained characteristic of all of Michotte’s research. (Below we will turn to another member of this school, Karl Btihler, who was the first to apply the school’s ‘method of systematic introspection’ to psycholinguistic issues). Michotte’s epistemology was neo-Kantian. He believed that the major categories of cognition csubstance, reality, causality) were innate, and the dominant direction of his work was to show that in origin these are innate perceptual categories. The immediate and compulsive impression of causality, for instance, arises under precisely &finable and quite restricted perceptual conditions. By inventive experimentation, Michotte could construct so-called ‘negative’ cases’ where thez:e perceptual conditions were noI fulfilled, but where experience would suggest causality; still, no impression of causality resulted. And even more convincingly, he set up ‘paradoxical cases’ where the perceptu& I conditions for causality were fulfilled, but in such a way as to contradict experience (so, for instance, if an object diminishes its speed at being hit by another object moving in the same direction); here he found his subjects spontaneou.sly and systematically reporting an impression of causality. As Michotte (1963, p. 220-221) put it, these cases provide a clear demonstration of the uselessness of any psychological theory which suggests that it is past experience which playsthe crucial part in setting up causal links’. Not surprisingly, Michotte had deep disagreements with Piaget, though they used to address one anoth::r as ‘le Maitre de Geneve and “le Maitre de Louvain’, and to understate their disagreements in highly polished language. When I read ‘The debate between Jean Piaget and Noam Chomsky’ (1980), SO beautifully edited by Piatelli-Palmarini, I had a strong experience of deja ~1. Exactly the same arguments had been going on betw 3n Piaget and Michotte more than twenty years earlier. When Chomsky w&es ‘The natural way to proceed, if we are trying to determine the nature of&, [the genetically determined initial state-W. L.] , is to try to find some property of the steady
D&i
vu?
189
state that is minimally affected by experience, a property for’which E (expetieme) is reduced as close to zera as possible’, he expresses exactly the logic of Michotte’s experimentation pith negative and paradoxical cases. The notion that important aspects of our behavior and experience art; bs sed on pre-given structure, over which we have little control, has been and I till is a Leitmotiv in my research. Additionally, there is the conviction that these basic structures have a modular organization with a maximum of urganization within a module, and a minimum of interaction between modules. This conviction I surely owe to my thesir supervisor John van de Geer who expressed it as the notion of ‘relative autonomy’ of subsystems. As a consequence, I dislike heterarchical theories in which all modules can ta1.kto i.11other mo dules; I prefer hierarchical organization in cognitive theory. Some examples can illustrate these theoretical starting points. In my work on binocular vision (Levelt, 1968), I have tried to determine the exjres’mteraction in brightness and rivalry. The findings leave no doubt that higher processes such as GestaIt formation or attention do not interfere to any substantial degree with the system’s activity: the cyclopean system is relatively autonomous, and the interaction between the eyes is fixed and rather simple, in full agreement with Hering’s (1862) nativist views on the visual system, and in contrast to Helmholtz’s empiricist position. In the same vein, we have shown that loudness interaction between the two ears is completely additive, i.e., that there is no interaction term at all: the two ears deliver their output without being affected by one another (Levelt, Riemersma and Bunt, 1972). Another example can be found in work on the perception of musical consonance I was involved in (Plomp and Levelt, 1965; Levelt, van de Geer and Plomp, 1966). There is no doubt that culture and experience are major determinants of consonance perception. Still, we couid show that, whatever culture has built, it is rooted in the given psychophysical structure .of the ear, more specifically in the so-called ‘critical band’ af pitch/loudness interaction. The situation is very similar to that of the perceptual origin of causality mentioned above: In both cases there is an immediate impression over which we have no control, but which may develop into an abstract cognitive category. The issue of relative autonomy is especially intriguing where perception
I9G
Willem J. M. Levelt
Hoer&, 1938, Schreuder, 1978) was th;at in verification no semantic decomposition takes place (for ‘true’ cases), and that the proLess is driven by the perceptual system without much feedback from the linguistic system: the most salient perceptual feature ‘looks for’ the most salient meaning component of the verb, and not inversely. Processing relations between perception and language has also become a major theme in our newly founded Max-Planck-Institute for Psycholinguistics. More specifically we study how speakers operate on spaCal representations in producing descriptions of spatial arrangements (such as living rooms, routes, or more abstract networks). A central concern here is what I called the speaker’s linearization problem (cf. LeveiP, 1981a, b): ? speaker will normally have to ‘linearize’ a spatial or other knowledge structure for expression. This requ.rement to determine an order of mention, and how to deal with it has been stud.ied by rhetoricians for millenia, but the issue is by and large ignored in modem psycholinguistics. ‘Wefound that there is a coherent set of linearization principles. They are, in part, determined by what is mutual knowledge in the speech situation, and in part by the spezkers’s working memory requirements. Neither of these are linguistic in nature, and we have growing evidence that linearization decisions are unaffected by formulation processes such as lexicalization and syntactic choice (Levelt and Maassen, 1981). Linearization, therefore, may have to be considered as a relatively autonomous procedure in the speaker’s formulation process. The same principles of linearization should hold whatever the modality of language (e.g., signed versus spoken, cf. Levelt, 1980). Given a chosen linearization, however, the speaker provides the !istener very systematically with linguistic cues which facilitate the reconstruction of the intended spatial or other complex representation They comprise, among others, anaphoric devices, modal expressions and deictical devices of various =fis. The Institute is deeply izvolved with the study of how such devices are used in both perception and production, and how they develop in first and second language learners. Some of the work is reported in Jarvella and Klein (1981). Mentioning our research on deixis gives me the occasion to return to Karl Biihler, Jvho spent the last 23 years ctf his life in total oblivion in America. The Jarvella and Klien book opens with their translation of Btihler’s highly original analysis of deixis in part II .of his Sprachtheorie ( 1934). Something else which seems to have been complet~ielyforgotten about Btihler is the fact that he moved psycholinguistics into the laboratoiry, something George Miller had to accomplish again half a century later. In 1908(a) Btihler published a study in which he measured comprehension latencies for complex sentences, and where subjects had to give introspections on their process of understan-
Dijti vu? 191
ding these sentences. Even Ericsson and Simon (1980) show no awareness that this was the first major study with verbal reports as experimental data. The systematic analysis of linguistic introspections, which is thus nothing new, has fascinated me, for both methodological and theoretical reasons ever since my first contacts with generative linguistics. As far as method is concerned the sharp dispute between Bahler (1980b) and Wundt (I 907) had made it clear that, whatever one’s theoretical position, one should take intuitive reports as data just as any other behavioral data. Michotte (1954) clearly took this position to consider ‘les rtiponses verbales tomme des reactifs differentiels’, as do Ericsson and Simon. What I have tried to add to this notion is that the interpretation of such introspective data requires (a) a theory of the subject mat& under study; in case of judgments of syntactic cohesion or grammaticality this can, for instance, be a generative grammar of the language, and (b) a measurement or interpretation theory, i.e., a theory about how the (intuitive) data relate to the entities in the theory on the subject matter. In Level& 1974, Vol. 111,I developed formal interpretation theories for judgments of synta,ctic cohesion, and used them to study difBrent generative accounts of English and Dutch. The coming decade may produce some more work along these lines. As far as a theory of linguistic intuitions is concerned the question is: where do they come from? It is often argued that grammaticality judgements require sema.ntic interpretation of the test sentence. A reaction time study (Levelt et ok., 1976) shows that this is not so: syntactic judgment can be a relatively autonomous process. We have, furthermore, started research on the causes and functions of linguistic awareness in children (Levelt, Sinclair and Jarvella, 1978). Nothing would have to be changed in current theories of language acquisition if children were to show no linguistic awareness at all. Are poccurrences of linguistic awareness in the child ‘indeed mere epiphenomena, or is the child’s ability to reflect on language o:ie of the innate predispositions which are indispensable for attaining coherent interaction between independently developed functioning procedures (cf. Karmiloff-Smith, 1981)? My feet are in the Old World’s psychology, my hands grope around in the New World’s cognitive science. The Max-Plan& cj;ociety’s establishment of an Institute for Psycholinguistics will surely contribute to making this tingling tension productive, not only for myself, but also for large numbers of young scientists who will be shaping our field for the decades to come. References JHihler, K.(19080)
Tatsacheqund Probleme zu einer Psychologie der DenkvorgiSnge.II. Uber Cedankenzusammenhtige. Archiv fir die gemnte Psychofogie,12, l-23.
Biihler, K. (11908b) Antwort auf die von erhobenen EinwBnde gegen die Methode der Selbstbeobachtung an experimentell erzeuflen Erlebnissen. Archiv $!.4rdie Gesamte Psychokxie, 12, 93-123. Biihler, K. (1934) Sprochtheorie, Jena, Fischer. Ericsson, K. A.., and Simon, H. A. (1980) Verbal. reports as data. P$ychol Rev., 87, 2115-251. Bering, E. (1862) Beitrgge xur physiolqgie, V. Vom binokularen Tiefszhen. Kritik einer Abhandltw van Helmholtz aber den Horopter. Leipzig, Engehnann. Heymans, G. (1’903) Das Kiinftige Jahrhundat der Psychologie. ln G. Heymans, Gesummeltekleinere &tiftea IL Den Haag, Martinus Nijhoff, 1927. Jarvella, ‘R. and Klein, W., (eds.), (1981) Speech, pi&e and action: Studiesin deixis and related topics Chichester, Wiley. Karmiloff-Smith, A. (1981) Using metalin&stic data for a process-oriented approach to language acquisition. Max-Planck-lnstitut fur Psycholinguistik. Manuscript submitted for publication. Level& W. J. M. (1968) On binoculmtivoiry. Den Haag, Mou~,-:~. Levelt, W. J. M. (1914) Forn&mmrxars in &zguisticsand psycholinguistics.3 Vol. Den Haag, Mouton. Level& W. J. M. (1980) On-line processing constraints on the properties of signed and spoken language. In U. Bellugi and M. Studdert-Kennedy, (eds.), Sign& and spoken langwge: Bio&gical constraintson EnguM? form. Weinheim, Chemie. Level&W. J. M. (1981s) The speakers’s linearization problem. Discussion Meeting on the Psychological Mechanisms of Language. Phil mar& Roy. Sot. Land. Level& W. J. M. (198lb) Linearization in describing spatial networks. ln 8. S. Peters and E. Saarinen, (eds.), &cesr?a beliefs, and questions: Essays on the semanticsof natural hnguage and hnguage processing. Dordrecht, Reidel. belt, W. J. M., Sinclair, A., and Jarvella, R., (1978) Causes and functions of linguistic awareness in language acquisition. In A. Sinclair, R. J. Jarveila, ! and W. J. M. Levelt, (eds.), 77re child’sconception of h7guclge. Heidelberg, Springer. Lcvelt, W. J. M. and Maassen, B. (1981) Lexical search and order of mention in sentence production. In W. Klein and W. J. M. Levelt (eds.), ctossing the boun&ries in linguistics Studies presented to Munfied Bi’swi9ch. Dordrecht, Reidel. Levelt, W. J. M., Riemersma, J. B., and Bunt, A. A. (1972) Binaurt! additivity of loudness. Brit. J. rMlk. slut fiychol, 25,51--68. Levelt, W. 1. M., Schreuder, R., and-Hoer&, E. (1978) Structure and use of verbs of motion. In R. N. Campbelland P. T. Smith, (eds.), Recent advancesin the psychology of language. London, Plenum Press. level& W. 1. M., Van de Geer, J. P., and Plomp, R. (1966) Triadic comparisons of musical intervals. Brit. J. nwth. stat.&vchoL, 19,163-179. lcvolt, W. 3. M., Van Gent, J. A. W. M., Haans, A. F. J., and Meijers, A. J. A. (1976) Grammaticality, paraphrase and imagery. In S. Greenbaum (ed.), Aaptabiiiiy in hnguuge. Den Haag, Monton. Michotte, A. (1954) Autob&graphie. Louvain,Nauwelaerts. Michotte, k (1%3) The perception of caw&y, London, Methluen. Piat&liPal~, M., (ed.), (1980) Language and learning. The debate between Jew &get and Nwm abmsky. Lox&ion,Routledge and KeganPaul. pfomp, R. and kvdt, W. J. M. (1965) Tonal consonance and critical bandwidth. J. ucoust. Sot., 38, 518-560. Schreuder, tL (1978) S&dies in p@zol&c&gy, Withspectizfrefmence to verbs of motion. Dissertation,, Nijmegen University. tindt, W. (1907) Uber Ausfrageoxperixnente turd tiber die Methoden zur Psychologie des Denkens. &W~~b.w&che SWdkm 3,301-360.
193
&&tion, 10 (1981) 193-196 @ Elsevier Sequoia S.A., Lausannr - Printed in The Neiherlands
Natural and unnatura! cognition ELIZABETH
F. LOFTUS+
University of Washington
About 8 years ago, I worked with a lawyer on a case involving a yoting woman who had killed her boy friend. The.prosecutor called it first-degree murder, but her lawyer claimed she had acted in self-defense. What was clear was that during an argument, the defendant ran to the bedroom, grabbed a gun, and shot her boyfriend six times. AL the trial, a dispute arose about the time that had elapsed between the gra.bbing of the gun and the first shot. The defendant and her sister said two seconds, while another witness-a friend of the deceased -said five minutes. Th.e exact amount of elapsed time ma& all the difference in the world to the defense, who insisted the killings had occurred suddenly, in fear, and without a moment’s hesitation. Who should be believed, the defense witnesses or those of the prosecution? How likely is it that a person would remember accurately the crucial elapsed ti.me? How good are people at perceiving and remembering the, details of complex experiences? These questions and others were raised by this killing, and the lawyers turned to psychology for some answers. When asked, I naively assumed that the psycholog:J of perception and memory would surely have something seful to say. But the answers were not to be found in the current textbooks of cognitive psychology. The literature on sensory memo,ry -showing that when a stimulus is terminated suddenly, some information persists for a brief period of timewas not useful. The products of intense research on the serial position curve were not useful. The mathematical models of rticognition memory provided not a clue. For one thing, most of the experiments had been deliberately designed to investigate constrained laboratory phenomena. An intei’est in the study of the loose and broad phenomena that had so excited Sir Frederick Bartlett was only just beginning to be revived. In short, despite the advances of cognitive psychology as a science, its applicability to testimony about natural events was minimal. Fortunately there was an old literattire +o w&h I could turn. In fact, one of the oldest areas of applied psychology in general was the study of the reliability of eyewitness testimony. At the turn of the century, a numberof *Reprint requests should be sent to Elizabeth F. Loftus, Depaztment of Psychology, Washington, Seattle, Washington 98195, U. S. A.
Univers.itY cf
194
k-. F. Loftus
Txperimenters designed their studies to simulate realistically the observation and reporting of events. In a typical scenario, an argument between two people erupts in the midst of a classful of unsuspecting witnesses. After the staged incident, the instructor solicits reports from the witnesses. Inaccurate recollection of almost every facet of the event--words spoken, clapsed time, and so on--was more common than not. This paradigm has been adapted by current researchers of eyewitness testimony, and much of their current work is aimed at discovering the conditions under wh.ich testimony about complex events is more or less accurate. It was in this liteirature that some useful tidbits of information could be found; for example, the fact that people tend to overestimate the duration of complex events, especially stressful ones. The field of cognitive psychology seems now to be shifting toward an increased interest in “natural cognition”. We have seen a growing interest in how people store and retrieve information from stories, conversations, and college lectures, as well as details about their own personal experiences. In this atmosphere, research on eyewitness testimony is flourishing. I have heard some of my colleagues refer to this research disdainfully as “mere applied psychology”, but I think they have missed a point. There does not exist a single continuum between theoretical and practical research; we are not IIOW observing a shift in a “practical” direction with a loss of interest in theoretical issues. Rather, a more useful conceptualization is one of a twu-by-two design: Any given piece of research can be characterized as having theoretical interest or not on the one hand and as having practical relevance or not on the other. Some research is both theoretically useful and has important practical significance. Other research may have theoretical interest, but no immediate practical importance, and the converse is also possible. And finally, some research (unfortunately) is not particula.rly useful either theoretically or practically. At the risk of immodesty, I’d like to believe that my own work on eyewitness testimony falls into the first cell, that is, it has some theoretical as well as practical usefulness. In this work, I hav,: shown that information about an event can be introduced into a person’s recollection after the event has passed. The post-event information can supplement the previously acquired memory as, for example? when a witness mistakenly remembers seeing a barn that never existed. Cir, it can actually transform memory, as when a witness comes to believe that a reci traffic signal was green. This research has practical usefulness in pointing out the ways that multiple witnesses to the same event, or a police interrogation itse2, can cause changes in a witness’s memory. However the research simultaneously bears on a number of theoretical questions; for example, questions concerning the nature of human forgetting. Does forget-
Natural and unnatural cognition
195
ting consist of an actual loss of stored information. x does it result from a of access to information, which, once stored, remains forever? My research, wbile providing no absolute answers, is at least suggestive of the idea that some memories may undergo alterations. One of my hopes with this research is to cause my colleagues to question the widespread belief in the permanence of memory and the especially shabby nature of the evidence used to support this belief. Even when a piece of work has theoretLa.l and practical utility, a questic;n still can be asked about its degree of usefulness. For example, despite the complex nature of the stimuli to which subjects of eyewitness research are exposed, there are still those who would argue that the results do not tell us much about the behavior of actual eyewitnesses in real life situations and thus have limited practical utility. Some argue that the inaccuracies in the simulated setting are larger than one would find in the real world where witnesees who offer testimony in court are self-selected to be those who really got 1 good look. In the simulation many subjects may have been napping. others would argue that the inaccuracies in the simulated setting are smaller than one would find in the real world where witnesses are likely to be extremely frightened. In the simuiation many subjects may have been rather bored. A further problem arises when one attempts to apply the psychological research to the behavior of an actual witness. In the experiments, the data consist of averages of the reports from an entire sample of subjects. In the legal world, it cannot be assumed that because eyewitnesses in general are fallible, the testimony of a particular eyewitness probably is fallible as well. At present, our data can only reveal factors that are likely to be potentially biasing. Cognitive scientists will continue to conduct basic research that has no immediate practical applicability, and as interest in practical issues grows, their work must not be discouraged. Too many’important discoveries m’ght never have been made without the intense efforts of behavioral scientists who were fueled by sheer intellectual curiosity, rather than by immediate practicality. For example, c,iespite the fast that the real world never allows us to have feedback about certain bodily functions, research on biofeedback has revealed important findings about our.ability to control these allegedly involuntary responses. Despite the fact that the real world rarely presents the human observer with 50 msec. flashes of kights or letters, research in this paradigm has revealed a “ sensory memory” that can persist beyond the termination of a stimulus. Despite the fact that the real world does not present chimparLees with the opportunity to acquire American sign language, research in person-made environments designed to enhance the acquisition of language indicates its viability. Although some ctilleagues will disagree with the value loss
196
E. F. Loftus
of the particular instances I have cited, they illustrate a point, namely, much can potentially be discovered in unnatural environments. Experimental conditions that faithfully reproduced the external world might never have detected these phenomena. Some fraction of investigations of unnatural cognitions will pave the way for some crucial developments of the future, just as the work on the growth in cell cultures of antigenic types of virus paved the way for the development of the Salk vaccine. In conclusion, I have suggested a scheme in which any piece of cognitive research might be characterized according to the extent to which it is theoretically useful, and, along a different dimension, the extent to which it has immediate practical utility. It is no longer necessary to argue that the “Most Likely To Succeed” prize should be awarded to those who conduct one type of research versus another. It is only research that has neither theoretical nor social utility that we might ask to walk the plank.
CogMmz, 10 (1981) 197-200 @Elsevier Sequoia S.A., Lausanne -
197
Printedin Tl?eNetherlands
Artif icisl intelligence - a new theroretical psychology? H. C. LONGUET-HIGGINS+ University of Sussex
The phrase “artificial intelligence” tends to arouse passions and prejudices which endanger rational discussion. It suggests that computing science and advanced automation have now reached the point at which we can start to construct machines with superhuman faculties, and this idea has been seriously entertained by a number of computer scientists. In this note 1 shall try to dispel both the prejudices and the passions, and to propose that artificial intelligence, whatevntr its merits or defects as a technological aspiration, can provide us with ways of thinking about the human mind which are of great potential value in the formulation of cognitive theories. It is necessary to begin by making two disclaimers. The first of these relates io the comparison often made between the human brain and the digital computer. In an obvious sense the brain -or rather the central nervous system as a whole-is the body’s computing system; it has the formidable task of collecting and collating a mass of co1lplex and detailed information from the outside world (and from inside the body) and controlling our movements and secretions in such a way as to meet our biological and social needs. But the structure of the central nervous system is profoundly different in almost all respects from that of a computer. Most existing computers possess just one central processor which carries out, one at a time, every single logical operation demanded by the computation; in the central nervous system, on the other hand, we have a vast array of processors, all jvorking in ;>aralleland only comparing notes when it is necessary to do so. Another vitally important difference between the brain-body system and any computer-controlled machine is the astonishing versatility and sensitivity of our eyes, ears and fingers; most of the sensors and effecton of existing robots are pathetically crude and clumsy by comparison. It would therefore be absurd to imagine that we can do justice to the information-processing abilities of humans-or even dogs-by a facile comparison with our own artefacts. The second disclaimer is of a different kind: it concerns the level of abstraction at which it seems appisopriate to compare our thoughts with the succw sive steps in a computer program. One of David Marr’s most important contributions to theoretical psychology was to distinguish between three levels uReprint requests should be sent to H. C. Lmguet-H&ins, Laboratory of Experimental Psychology, Universityof Sussex, Brighton BNl 9QG, %@and.
1%
H. C. Longuet-Higgins
of abstraction at which one might discuss a given cognitive skill. At the highest level one wilI need to examine the nature of the task itself, with a view to d&ovefing the constraints which the nature of the task imposes upon alternative methods of accomplishing it. An analysis of this sort will be equally relevant whether the task is to be carried out by a human being or a machine. At the next level, having adopted a particular method of solution, one will need to :;pecify one or more effective procedures, or algorithms, for proceeding from the initial to the final state; such algorithms can most easily be specified as computer programs written in a high-level language such as ALGOL or LISP., FinalIy, if one is interested in the underlying physiology, one will need to make specific proposals about the way in which the successive steps in the program might be implemented in neural tissue; but what directly concerns the cognitive psychologist is the logic of mental processes, not their neurophysiologicsl correlates. The analogy between the physical processes taking place in a computer and those which occur inside someone’s head is likely to be remote in the extreme; and this is my second disclaimer. But what justification might there be for attempting to describe the processes of human thought and perception in algorithmic terms? What reasons might there be for thinking that we can? Everything hinges, I suggest, on what we take to be the goal of cognitive psychology. What the psychologist would like to do, surely, is to describe in logical terms the psychological events which mediate between our experiences and our actions. The question is: how would we recognise such an account if we were offered one? To this question artificial intelligence proposes the challenging answer: we should recognise an interpretation of some cognitive skiIl as a fully e:xplicit theory if and only if it could in principle be used for the construction of an automaton which would simulate that skill. Such a criterion might well seem altogether too exacting, by making it virtually impossiblce to construct a fully explicit theory of any cognitive skill whatever. The idea of a fully explicit theory is, nevertheless, a valuable one, in that it induces a healthy dissatisfaction with cognitive theories which are avoidably imprecise or inattentive to detail. At this point the question naturally arises: if and when someone advances a fully explicit theory of some cognitive skill-perhaps embodiell in a computer program, perhaps not- how is one to tell whether the theory is right or wrong? The-ab a3, ;n unfortunately, no easy answer to this question; but there are a number of relevant observations to be made. First, the process of constructing a fully explicit theory, in the form of an effective procedure, is itself illuminating; it invariably draws attention to important questions of detail which might otherwise have been overlooked. To take an example, the history of transformational grammar is littered with discarded hypotheses
Artificial intelligence-a
new theoretical psychology ?
199
bearing witness to the progress of the subject; only a hypothesis which is suffici&!*yr explicit to be possibly wrong is of any value in the progress of a science, and it is greatly to the credit of modern theoretical linguists that they have preferred a hazardous precision to a comfortable vagueness about matters of detail. Secondly, all that one can reaso,lably expect of a scientific theory is that it shall be clear, plausible and suggestive of new experiments or observations, and, of course, that its predictions shall be firmly connected with its assumptions. Only then, if some of its predictions are eventually falsified, can one be sure that at least one of its premises was at fault, and can attempt to track down the offending assumption. To repeat: the particular contribution that artificial intelligence can make to psychology is the concept of an effective procedure for the performance of a cognitive task. The concept of an effective procedure is general enough to cover such diverse cognitive skills as the acquisition, production or comprehension of language, the perception of speech “r;s music and the mental reconstruction of the visual world. I mention these particular skills because they have actually been modelled, with varying degrees of sophistication, by workers in artificial intelligence, and are tasks which human beings can actually perform reliably and reproducibly. And if one is able to discover an effective procedure for the accomplishment of a given task, then one can check whether or not it really is effective by translating it into a suitable ;3rogramming language and running the program on a computer. One almost always finds, in real life, that the first few editions of a program are faulty in a number of respects: either the procedure fails to specify what should be done in one or more exceptional situations, or the processes to which it actually gives rise are not at all as planned, perhaps because one has overlooked some essential feature of the input or of the relevant cognitive states. When an AI program of any interest eventually reaclres the journals, one can be quite sure that i0 has already been subjected to an extensive and ruthless series of revisions; if the final product seems unsatisfactory, that ma.y be because constructing good theories is at least as difticult as designing good experiments. One might compare the very short history of artificial intelligence with the very long history of alchemy, the ancestor of modem chemistry. The alchemists’ set themselves the task of tinding elixirs which would turn base metals into gold, or prolong life indefinitely; modem chemistry has replaced such fantasies with a much deeper understanding of matter and its transfermations. ArtificiaI intelligence began with the ambition of constructing artificial people, but in its maturity is providing us with a new way of constructing theories of ‘huiiian cognition. The notion of an effective procedure is important in this connection for a number of reasons. First, it is a sufficiently
200
H. C. Longuet-Higgins
abstract, though precise, idea to be applied to informational processes of vird we must be prepared to deal with such tually unlimited complexity-an processes in studying human perception and cognition. Secondly, the descrip tion of cognitive processes as effective procedures does not commit one to any hypotheses about the physical nature of the central nervous system, and frees the psychologist to discuss the inner logic of mental processes independently of their neurophysiological correlates. Thirdly, a truly effective procedure can, always in principle, and usually in practice, be embodied in a computer program written in an appropriate high-level computer language. Computer languages are designed for just this purpose: to enable us to write down with complete clarity and precision any conceivable sequence of logical processes. Fourthly, the exercise of expressing a procedure as a program reveals quite mercilessly any logical deficiencies in the procedure. And finally, when the program is actually run on a computer, the output will almost always surprise its author in one way or another. Such surprises are a source of new ideas. If they arise from gaps in the logic of the procedure itself, they usually point the way to a new and better theory, founded on more interesting or plausible premises. It ic, perhaps, time that the title “artificial intelligence” were replaced by something more modest and less provisional. Various alternatives have been offered at one time or another; “experimental philosophy”, harking back to the seventeenth century, and “epistemics”, suggesting a more practical pursuit than mere epistemology. Might one suggest, with due deference to the psychological community, that “theoretical psychology” is really the right heading under which to classify artificial intelligence studies of perception and cognition ? This suggestion seems to accord closely with the relation which commonly holds between the theoretical and the experimental branches of a science. The task of the theoretician is to formulate hypotheses and to elicit their lo&a). implications as carefully as he can, with due attention to matters of internal consistency and predictive power; the duty of the experimenter is to coni’ront the predictions of a theory with firm and relevant observations, and to suggest points at which the theory needs modifying in order to bring it into line with experiment -if that is indeed possible. The time has now come, it seems, when the task of theory construction is altogether too intricate to be consigned to spare moments away from the laboratory; it is at Ieast as much of a discipline as good experimentation, and one is gratified to see that artificial intelligence is rapidly becoming a standard component of undergraduate courses in psychology.
Cognition, 10 (1981) 201-208 @ Elsevier Sequoia S.A., Liausanne- Printed in The Netherlands
201
Action, thought and language DAVID McNEILL” Uniwsity
of Chicago
Psycholinguistics, wedged in between two Tleighbors, has always depended on psychology ‘and linguistics in various ways. From psychology it has taken its methodology, especially the use of experimental methods, and a taste for information processing models. From linguistics it has taken its definition of theoretical issues and indeed of the field itself. Though psycholinguistics has absorbed elements from the juxtaposition of psycholrgy arJ linguistics, there has not been a true synthesis. There is a mixture of elements, a bit from psych Aogy, a bit from linguistics, but no rationale for this particular mixture and no reason for preferring it over some other. I believe, however, that a distinctive conception of language is possible, a truly psycholinguistic approach. To see this spy Teach, we must consider the individual and his use of lznguage. Each ocellrrence or token of a given language structure has its own in,. dividual life history. Someone says a sentence. It has emerged on this occasan from a confluence witbin the speaker from many sources. Someone else understands the sentence. This too results from a cozlfluen ;e from many sources within the hearer. PsycholSnguistics and linguistics are fields which apply to successive slices along this life history of given language structures. Psycholinguistics focuses on the language structure as it emerges during acts of speaking and listening by the individual, a’nd for psycholinguistics language struttuies have an emergent, dynamic, temporal dimen:.ion. Linguistics focuses on the finished language structure, after it has been created or underst’ood, and for it language structures are complete, static and instantaneous. In a psycholinguistic approach, language structures correspond to forms of human activity, Language structures are the result of things done by inclividuals. Sentences, phrases and words (as well w other stretches of speech) define actions, or performances. In linguistics, sentences, phrases and words ,are regarded as static objects, like crystals. Language structure regarded as an object is how language activities appear to view after they are *I wish to acknowledge the fmancial support of the National Institute of Mental Health and the I am grateful to Nobuko B. McNeiUa;ld Michael Silverstein for comments on the manuscript. Reprint requests should be sent to David McNeill, Department sf Behavioral Sciences, ,UniversiSyof Chicago,5848 South University Avenue, Chicago, Illinois 60637, U.S.A.
SpencerFoundation.
202
David Mc.Neili
over. The linguistic approach was defined by Saussure (1959 [ 19 15 I ) as synchronic and paradigmatic. A language is considered as a whole with all its parts, in relation to each other at a single instant of time; that is the meaning of the terms paradigmatic and synchronic. This approach of Saussure’s I see as having reached a kind of climax in the development of generative grammar Saussure spoke mainly of overt (surface) linguistic elements organized into mutually defined paradigmatic sets ‘of relations. A fundamenta.lly new contribution froIm generative grammar was to include in the paradigmatic set invisible, deducible only ‘abstract’ linguistic elements. (Who& 1945, introduced covert paradigmatic linguistic elements, which he termed cryptotypes, but apprer’c:; did not carry the idea further.) These elements were called in 1965 deep structures. With this new concept:ion, the synchsonic pa.tteming of language was extended to include invisible elements. It seems to me that the S:l;assurian tradition cannot be pushed much farther without new insights comparable in depth and importance to the discovery of abstract paradigmatic rbments. The current tlisputes over whether there are or are not transformations appear to involve minor descriptive matters by comparison. To define a proper role for psycholinguistics in the study of language, one pertaining to language regarded dynamically, differentiation from the modern form of Saussure’s synchronic conception is crucial. In the definition of the field which I favor, psycholinguistics is not busy merely with confirming the psychological reality of constructs proposed in synchronic linguistic descrip tions, but must formulate a new theory of language activity Language activities yield linguistic objects, but do not necessarily directly incorporate them. I am thus involved in an approach to language structure which takes an activity point of view. In this I chiefly follow the lead of Vygotsky ( 1962, 1978), who long ago formulated psychology in general as the study of human activity. i will now sketch some parts of what I think an activitv approach to language structure could be. To make this approach intelligible requires coordinating a number of topics. These fall into two classes which I describe in parallel in this paper. First is a new arrangement of observations designed to bring out the activity aspects of language; gestures, metaphors and the systematic comparison of thes: to speech. Second are the inferred structures of language activity itself; the structure of action and the relati,onship of action (which is synthesizing) to speech (which is analytic); images and the schemas which produce them; and the use of these image-producing schemas to present other concepts. Language activities can be subdivided into smaller parts only up to a certain point. The unit of activity cannot be broken down further. While it is possible to fiid and isolate in the language activity separate structural, functional,
Action, thought and langmge
203
motivational and actional elements, doing so results in the destruction of the activity itself (Zinchenko, 1980). What remains is a list I:..fparts, .not a unified activity. A unit of language activity contains elements of various types which have in common that they are acted upon together. I have been particularly interested in the role of manual actions in language activities, especially gestures, and the evidence they provide of concrete imagery during language acts. Gestures of a type I will call iconic are evidence that speakers create, concurrent with speech, concrete depictions of meaning. Iconic gestures offer a second channel of observation of the speaker’s mental representat:ions (speech being the first channel). Such gestures are expressions in action of images which exist simultaneously with the meanings being expressed li.nguistically. Kant ( 1973 [ 178 1 ] ) defined a schema as a procedure for producing an image of a concept. Iconic gestures appear to be images of concepts, and imply the existence of schemas which produce them. These schemas are, I lbdieve, fundamental elements in language activities which can be seen as playing a role in speaking itself. Examples of iconic gestures are the following from a child: Speech
,they urn wanted to get,,where Anansi was,
Gesture
BH in midair, move together L tG R
LH goes to rest
RH goes up and index finger extends, and makes arc in the air
The first gesture, co-occurring with ‘they urn wanted to get’, depicts pursuit without contact; the second gestar e, co-occurring with ‘where Anansi was’, depicts an enclosure. Pursuit without contact and enclosure are suggested as the i_lnagesused on this occasion to portray the speaker’s meaning, or concepts; these concepts were, in this case, the effort to reach, but the inaccessibility of, the character called Anansi, and the locus of Anansi at this point in the narrative inside of a fish (Anansi is a spider; the referents of “they’ are Anansi’s six sons; and the story itself is based on a folk tale of the Ashanti people of G.hana). Thus the images of pursuit without contact and enclosure were connected by the speaker to the concepts of effort-plus-inaccessibility and entombment, and imply the presence of schemas which produced of these concepts the images depicted by the two gestures. The concrete depictions of meaning enacted in iconic gestures are produced (a.ccording to the Kantian formula) by schemas, or procedures. Why should
204
LkvidIucAkill
these procedures operate during language activities? I am very much intrigued by the idea that they are., in some way, a fundamental part of the linkage of the action of speech production with thought and meaning structures. During the production of speech, speakers are able to control their motor movements (various articulators) by something seemingly incommensurable, m.eaning decisicns. Meanings from different sources (different lexical items) must be brought together and into contact with action (speech articulation). Procedures, or schemas, are, I propose, the meeting ground of meanings with action. This proposal entails that actions carry with them meanings; and iconic gestures demonstrate this fact. The proposal also leads to the idea that meaningful actions, those performed on objects in the world, underlie actions in other modes, those performed with the speech articulators. Image producing schemas can also be seen to play a role here. Schemas which produce images of concepts I equate with the type of mental representation termed sensory-motor by Piaget ( 1952) and actional by Werner (1948). Such sensory-motor, or actional, schemas represent meanings (concepts) which have the potentizl of being displayed in actions. Sensory-motor schemas are simultaneously actional and meaningful (i.e., are representations of things other than themselves). This property is perhaps the fundamental form of a mingling of meaning with motor action (Vygotsky, 1962) and can Be regarded as underlying the control of all speech articulation. Motor actions., even simple movements, are complex and hierarchically structured (Bernstein, 1967; Greene, 1972; Turvey, 1975). Only at upper levels of action is there consciousness of effect and intentionality. A person intends to pick up a pencil and is conscious of the effect of the action he performs. This intentionality and consciousness are parts of the action at the highest level, and it is at this level, therefore, where the person is conscious of the effect of his actions on the world, that actions can be organized into schemas which are able to represent things other than themselves. At lower levels of the motor control hierarchy are various synergisms which regulate the details of action performances as physical processes; force, velocity, timing, the selection of specific effecters, as wellas the moment-by-moment monitoring and changing of movements_ Because of the hierarchical structure of motor control, lower nodes can be substituted for by other lower nodes, without disrupting the higher levels of the hierarchy. Flexibility of motor pefiormance depends on this type of substitution. Actions can shift from hand to hand, for example, without the relation of the action as a whole to the individual’s goals and consciousness of effect changing. In the production of speech the same principle of substitution of lower nodes in motor control hierarchies perhaps applies. Into a hierarchy originally
Action, thought and ian!guuge
205
established to control manual movements, for example, lower nodes can be substituted which control speech articulator movements. Thus, deeply embedded within the speech process can be manual actions and the schemas of representation which they support. To initiate the production of speech, according to this model, the speaker must find sensory-motor schemas with which to represent his intended mean~gs (concepts) which may be qu,ite abstract. It is for this reason, I suggest, that schemas, or procedures, for producing images of a kind depicted in action are involved in language activities. That schemas for producing action images of concepts are a fundamental part of language activities is suggested by the depiction of images when the concepts in question are highly abstract. Iconic gesture?: dre no lcsv common with abstract meanings, yet in these cases there can be no question that the speaker has created images of the concepts. The following are examples from a discussion between two professional mathematicians: Speech
,and take its dual,
Gesture
LH, palm down, rotates so that palm is up a finite quotient,that factors through,
Speech Gesture
RH loops down and to the right in large sweep
From the concept of a mathematical dual, the speaker generated an image of something flipping over in space. From the concept of factorization through something he generated an image of something sweeping through space. These images imply schemas in terms of which the speaker presented t:ze m;thematical concepts that were his intended meanings. The schemas which generate images of concepts thus often embody metaphors (understanding one thing in terms of another). ?‘he gestures above can be interpreted as revealing metaphors. These are not hrerary metaphors, but metaphors built intoordinary language and used inday-today understanding of the physical and social world. Thus the idea of effort-plus-inaccessibility can be understood in terms of a metaphor (as in a common form of frustration dream) based on pursuit without contact. The gesture was an image which depicted the metaphoric presentation of the concept. The idea of a mathematical dual can be understood, in part, in terms Df a metaphor based on physical rotation of an object through space. Lakoff and Johnson (1980) have catalogued a number of metaphors of this kind. Many if not all are easily
206
David McNeil1
enacted in gesture form. That is because the metaphors use for their vehicles (Richards, 1936), or secondary subjects (Black, 1962, 1979), the concrete manipulation of objects. An important metaphor for pres:f:nting the idea of language- _itself is called by Reddy (1979) the conduit metaphor. This metaphor is the source of maqy gestures. According to the conduit metaphor, language is a container (e.g., ‘words are fuii of meaning’), meaning is a substance (e.g., “lay it all out on the table’), meaning is put into and taken out of containers (e.g., ‘putting your ideas into words ‘, ‘digging out the meaning’), and communication of meaning is the sending of containers full of meaning-substance over a conduit’ to a destination (e.g., ‘had a hard time getting through to him’). All of these metaphors are based on manipulations of objects and produce images which gestures readily depict. For example, a common gestural accompaniment of verbs of saying, ‘I was telling, explaining, asking, etc.’ is made with the hand extended outward, the palm up. the fingers extended, separated and slightly curved--holding, in fact, a container, the image of the concept which is denoted by the verb. Xnvolved in language activities, then, are metaphors, or schemas used to present something else, and these are based in many cases on the manipulation of objects. In this sense, manual actions are a deeper element of language activities than gestures alone, for they are the basis of the metaphors used in presenting the speaker’s meanangs. Combined with the speech channel, gestures give a rich basis, beyond the basis usually considered in linguistics investigations, for interpreting language activities. The gesture channel makes evident that speakers create, concurrent with speech, concrete images of meaning, arising in interesting cases from metaphors. Gestures, however, depict meanings on a quite different principle from the verbal channel, and this difference also implies that a different princiiple is involved in the underlyling metaphors and schemas. lconic gestures CZI be effectively instantaneous (even if stretched out in time they represent everything that they represent at once), whereas in speech meaning units are segmented into words and distributed across time. The gesture which accompanied ‘they urn wanted to get’, for example, represented at a single (effective) instant everything (and more) that in the speech channel was divided into four words and distributed across time. Vygotsky (1962) argued that words are the minimal units of language activity. They are the smallest parts of the speaker’s activity, dividing this activity and distributing it across time. In this respect words differ from subword units, such as morphemes. Decomposition of complex words such as ‘unfriendly’ into morphemes destroys the cormection of the word with any form of language activity; ‘un-‘, by itself, has only an abstract meaning of negation, whereas combined into a word the prefix
Action, thixlght and language
207
takes on life and meaning at the level of the activity. It is easy to imagine the type of language activity in which the word ‘unfriendly’ might appear, but impossible to imagine any activity in which ‘un-’ appears (without also specifying the rest of the word it combines with). Words have a particular kind of motivation in psycholinguistics, therefore, which object-oriented constructs such as morphemes lack. Though words are often recognized as basic linguistic elements, they reflect an ability not so much to combine elements into larger structures, as an ability to divide language activities into minimal segments. The most fundamental lesson we can learn from the combination of the gesture and speech channels is that in language activities there is an irrierplay between two quite distinct modes of meaning representation-the instantaneous and unsegmented (images and metaphoric schemas)and the segmented and successive (words and grammar). The non-reduction of language activities beyond a certain point forbids separating these modes from each other without destroying the language activity itself. In an activity approach to language structure-that is, in the psycholinguistics I am advocating--it is impossible to consider the traditional components of langllage structure (lexicon, grammar) in isolation from the unsegmented wholistic representations of meaning shown in gestures, images and metaphors. The psycholinguistic activity oriented approach I have sketched is meant to work in’rc,a synthesis with the object oriented approach to language structure pursued in synchronic linguistic descriptions. This synthesis would provide, in a comprehensive manner, coverage of the life history of given language structures by referring to successive slices of this life history. The result is a rationalized division of effort between adjacent fields which differ in their approach to language. The psycholinguistic approach I have presented is discussed at greater length in several other places; chiefly, a paper published in 1975 and two books, one published in 1979 and one to be published in 1982. The latter book treats in some detdil what for want of space I have omitted completely from this sketch, namely, the shaping of language structures for contextual and discourse functions. The ideas of the activity approach have been evolving continuously in these publications, though the basic approach has remained the same. References Bernstein, N. A. (1967) The Co-ordinationand Regulationof Movements.Oxford, Pergamon. Black, M. (1962) Models and Metaphors:Studies in Lang-rageandPhilosophy. Ithaca, Cornell University Press.
208
David McNeil1
Black, W. (1979) More About Metaphor. In A. Ortony (ed.), Metaphor and Thought. Cambridge, Cambridge University Press. Greene, P. (1972) Problems of Organization 0;’ Motor Sys&:ms. In R. Rosen and F. M. Snell (eds.), i+o@ss in Theoretical Biology PJ&. 2). New York, Academic Press, Kant, I. (t 973) critique of &we Reason. London, Macmillan. Lakoff,. G., and lohnson, M. (1980) &%&@zo~sWe Live By. Chicago, University of CXcago Press. McNeil&D. (1975) Semiotic Extendon. In R. Solso ted.), dnformrion Processing and Cognition. Hillsdale, NJ., Erlbaum. McNeil&D. (1979) Tdre CbnceprtutJBusisofLunguage. Hillsdale, NJ., Erlbaum. McNeill, D. (In press) Action, Thought and. Language: A New Approach to Psycholinguistics.New York, Harper & Row. Piaget, J. (1952) 7Ise &it& ofhtrelligence in Bildren. New York, International Universities Press. Reddy, M. (1979) The Conduit Metaphor -A Case of Frame Conflict in Our Language about Language. In A. Ortony (ed.),Meruphclrand ZItoughr.Cambridge, Cambridge University Press. Richards, I. A. (1936) The Philowphy of Rhetoric. New York, Oxford University Press. Saussure, F. de (1959) c;oUrse in General Linguistics.New York, Philosophical Library. Turvey, M. T. (1975) Preliminaries to a theory of action with referencles to vision. In StatusReport on Speech Research (Jan.-Mm&t,1975). New Haven, Haskins Laboratory, SR-41. Vygotsky, L. S. (1962) nought aaIdLanguage. Cambridge, Mass., MIT Press. Vygotsky, L. S. (1978) Mindin Socrety: T&eDev4ogment of HigherPsychologicalProcesses.Cambridge, Mass., Harvard University Press. Werner, H. (1948) Cbmpamrive Bychology of Men&l Developmenr. New York, Science Editions. Wharf, B. IL. (1945) Grammatical categories, f,unguuge, 21, l-l 1. (Reprinted in J. B. CarraIl {cd.), tingzurge. 7bughr and R&i@?, Cambridge, Mass., MIT Press, 1956.) Zinchenko., V. P. (1980) The ideas of L. S. Vygotsky about units in the analysis of mind. Papergiven in Ckto’berat a conference, Culture, Communication, and Cognition: VygotskianPerspectives, at the Center fo: Psychosocial Studies, Chicago.
Cbgnition, 10 (1981) 209-214 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
209
Lexical access:A perspectivefrom pathology JOHN C. MARSHALL* FREDA NEWCOMBE The Radcliffe Infirmary
The research program o I* which we (and many other neuropsychologists) have been working for the last decade concerns the nature of the interface between perception and the language faculty. The specific problem that has attracted much of our own attention relates to the ways in which orthographies mediate between vision and lexical retrieval; the prims.ry assay that has been employed is the study of ‘double-dissociations’ (Teuber, 1955) in the residual reading skills of adults who, prior to sustaining brain-injury, were fully literate. The c. igina! goal of Che i3trJgram was to construct a rational classification of the acquired dys!e.:;*.i; (Marshall and Newcombe, 1973). That is, we sought a systematic ar.ai,4s of b:-eakdown that mapped onto a quasi-formal account of the funct@i;til components required in normal reading [Morton, 1979) and that had a fairly direct relationship to some of the strategies encouraged in the acquisition of reading skills (Newcombe and Marshall, in press). The explanatory mode that we adopted was traditional ‘diagram-making”’ (Lichtheim, 1885) Boxes and arrows were employed as a notation for computational centers (or the representations that they assigned) and the information-transmission route; whereby they are interconnected. The behaviour of patients with (often relatively focal) brain-damage was then interpreted in terms of hypothr,!ical ‘lesions’ of the boxes and arrows postulated in the theoretical model Adoption of this research strateg: , analogous to the use of mutations i.n the identification of genes, and hence of control points in a biological system, quickly led to a new classification of the acquired dyslexias that could, at very least, be neatly summarized by relatively simple block-diagrams (Coltheart, Patterson and Marshall, 1980). The primary symptom-complexes of the new taxonomy include deep dyslexia (Marshall and Newcombe, 1966), surface dyslexia (Marshall and Newcombe, 1973), phonological dyslexia Beauvois and D&ouesnC, 1979), direct dyslexia (Schwartz, Saffran, and Marin, 1980), semantic access dyslexia (Warrington and Shallice, 1979). and word-form dysfexia (Wa.rrington and Shallice, 1980). Althougn there are many substantive arguments that still remain to be resolved within the notational conventions of simple informstion-processing models (see Morton, *Requests;or reprints should be addressed to J. C. Marshal& Neuropsychology Unit, Neuroscience Group, The Radcliffe Infirmary, Oxford OX2 6HE, England.
2 10
J. C Marshalland F. Newcombe
page 227), it is our current belief that this work is rapidly reaching the stage where the ad hoc proliferation of boxes and arrows will weaken the explanatory power of the approach to the point of vacuity. The reader who is familiar with the history of neurolinguistics will recall that this danger is not without precedent (Moutier, 1908). We accordingly outline in this paper a reconstruction of the original program that will, we hope, provide a framework within which some unifying principles may eventually be found. P;e-theoretically, we assume the notion ‘word’ and ask the question: Given an orthographic string of word-length, how does that string access the semantic, syntactic,. and phonological inforrnation that is arrayed against it in thz competence of a mature reader? Following Marr ( 1980), we further assume that the system so isolated must be understood in terms of four quasiindependent sub-theories. The fnst of these is the theory ot tne orthography itself. Most children learn to read after the core structure of the language, including much of the core vocabulary, has been triggered by auditory/vocal experience; it accordingly makes sense to regard the theory of written language as a statement of the mapping between the form of an orthography and a (set of) level(s) of representation made available by virtue of having acquired the spoken language (Klima, 1972). To a first approximation, extant orthographies can be divided into alphabets, morphophonemic (English) or phonemic-phonetic (SerboCroatian), syllabaries (Japanese Kana), and wordor morpheme-based scripts (Japanese Kanji) according to the smallest units at which correspondences between the written and the spoken language can be expressed. The second sub-theory would specify which of the mappings that the orthography in principle makes available are actually utilized under different conditions, and would provide an account of the algorithms that effect the mappings. Let us illustrate this with a hypothetical example. Consider a shallow orthography such as the international Phonetic Alphabet (narrow transcription). Words written in IPA could be read according to a strategy that segmented the orthographic string into letters and mapped those letters, one at a time from left-to-right, into a phonetic code that was both the representation that triggered articulation (sub.ject to so-articulation constraints) and the access code for retrieval of syntactic and semantic information pertaining to the word. The orthography ullou~ such an algorithm; it does not, however, demand it. Thus it is equally possible in point of logic that an IPA letter-string could be recognized as a stored word, by application, say, of Forstzr’s decision-tree algorithm (Forster, 1978); the word-address computed by the algorithm could then call dire&” the phonological and syntacticosemantic information associated with that a,ddress. In short, the structure of an orthography imposes constraints on possible algorithms (there is no pro-
Lexical Access: A perspective from pathology
211
cedure: whereby the individual strokes of a Kanji character could be mapped into phonological elements) but it does not uniquely determine a processing algorithm (Marsh&, 1976). Further constraints on the nature of the algorithms that are actually used can be obtained by study of the stimulus-variables that affect either rate of access in normal subjects or yrobcbility of correct access in brain-damaged patients. One such variable is word-frequency, or perhaps the highly-correlated variable of age-of-acquisition. Thus in lexical decision tasks, the speed at which orthographically-permissible tokens can be assigned to the types word or non-word iis, for the word tokens, an inverse function of the frequency with which the tokens appear in a (large) sample of text. One possible algorithm that would capture the constraint can be !informahy) stated as follows: Take the first n elements of a presented letter-string and use them to address directly all the words of the language that begin with those n elements; order by frequency the subset of the vocabulary so selected and evaluate each item in turn against the full input letter-string; accept the first item that matches the fumlinput. An algorithm of (approximately) this nature is presented by Taft (1979). An alternative proposal might involve the Iollowing steps: Assign to each stored word-representation a number that is inversely correlated with the frequency of occurrence of the word in the language;. quantize stimulus-nformation in such a way that numerical weights are associated with ‘fef;tures’ OC the stimulus, where a feature is some visual element or property of the sthmulus array; transmit the quantized and weighted features of the stimulus in parallel to all stored word representations; when a stimulus-feature matches a stored feature of a word add its numerical value to all words containing the stored feature. Continue this process until the sum of such matches reaches the originally-assigned value of a word, An algorithm of (approximately) this nature underlies the models presented by Morton ( 1979). Construction of an appropriate algorithm for word recognition must in turr. be distinguished from an account of the particular mechanisms that realize the algorithm(s). A third sub-theory will specify the nature of the storage devices, adders, multipliers, transducers, feedback loops and so forth that are required to implement the algorithms. Thus one interpretation of Morton’s model, tile interpretation preferred by Morton (1979), req- ‘-es ‘passive’ adders (or multipliers) in order that the ‘level of activation’ of Jred units can be incremented until their ‘threshold’ is reached; it also reqk:ires some forms of transduction into a ‘neutr 11’code in order that contextual (syntactico-semantic) information can summate wl+h stimulus information in a particular modality. The irlodel proposed by Taft (1979~ requires a short-term storage device that can hold stimulus information until the members the
of
2 I2
J. C. Marshalland F. Newcombe
candidate recognition list can be fed back for evaluation; the evaluation process itself presumably requires some kind of ‘comparator’, the nature of which must be specified. The algorithm for direct conversion between a letter-string and its phonological representation (illustrated in simplified form in the IPA example and believed to be implicated in the performance of patients with ‘surface dyslexia’) requires some kind of scanning device for segmenting the input string and fee&:; it(: elements sequentially to a ‘look-up’ table for the assignment of grapheme-phoneme correspondences. The mechanisms that realize an algorithm are the first theoretical entities for which it makes sense to raise the issue of neuronal representation. Finally then, a fourth sub-theory will provide an account of the neuronal hardware that instantiates the mechanisms. The relevant hardware is deduced, in the first place, from clinico-pathological correlation. Thus necropsy studies show that relatively ‘pure’ disorders ijf written language processing are associated with pathology centered on the junction of the occipital, parietal, and temporal lobes of the dominant (.typically left) hemisphere. Functional (or at least symptomatic) considerations thus suggest that ‘modular’ areas should be found in this region. If we defme areas anatomically, that , :$ in terms of a well defmed specialized cytoarchitecture and well defined ingut and output connections it does indeed seem that distirlct areas can be isolated within the broad region of pathology that can be demonstrated by histology at necropsy (Braak, 1980; Galaburda and Sanides, 1980). In some ins:ances , classical clinico-pathological studies have provided quite tight correlations between sym]ptom-eomplexzs of acquired dyslexia and focal damage. Thus dyslexia without dysgraihia is frequently seen with combined lesion of left calcarine cortex and the splenium of the corpus callosum (Geschwind, 1962), and with ‘subangular’ lesions located deep in the white : natter below the angular gyrus (Greenblatt, 1976); dysphasia with dysgraphia is often seen with pathology centered on the angular gyrus itself (Geschwind, 1962). One must note, however, that lesion sites only make ‘functional sense if the detailed character of the patient’s impaired and preserved skills can be interpreted in terms of lccalized mechanidns and their interaction. Thus in the taxonomy proposed in Newcombe and Marshall (19s 1), dyslexia with-and-without dysgraphia are no longer unitary syndromes but rnusf rather be fractio.nated into a variety of distinct conditions. Our general line of argument, then, follows Marr (1980) in claiming that the capacities of’ any complex biological system should be studied at all four levels of interpretation. We also assume that ‘bridge-laws’ must be formulated to link the levels. We accordingly see no virtue to the suggestion that theories formulated at one level should be impervious to criticism derived from data
Cexkal Access: A perspective from pathology
2 13
obtained at another level. What one should seek to maximize is the power of the theory as a whole; if this involves reinterpreting or overthrowing a claim about algorithms or mechanisms on the basis of what is known about basic component and circuit analysis for neuronal elements, then so be it. Some of the serial algorithms that have been proposed (admittedly somewhat tonguein-cheek) for word-recognition (Goldiamond and Hawkins, 1958) can be ruled out on the basis of the fact that speed of nervous conduction precludes their operation in real-time. Conversely, the requirement that a plausible mechanism (and algorithm) be ;;*vailablemight give one pause before accepting a particular anatomical interpretation of lesion data. Many ‘disconnectionist’ analyses that demand callosal crossing thus fail to specify the nature of the encoding and decoding machinery that would be needed to effect the requisite transfer of information. And, more seriously, one might expect that white matter lesions (within or between hemispheres) may well affect the computational capacities of their destinations over and above the role that long fibre tracts play in information-transmission. We will conclude with. a brief summary of why we believe that a partitioning into four levels of theoretical interpretation is a prerequisite of further advances in the neurolinguistics of reading (and neuropsychology more generally). A concern with the nature of orthcgraphies should draw attention to the necessity of studying a much wider variety of structural types than is currently under consideration (but see Sasanuma, 1980); we want to distinguish between ‘core’ reading skills that are needed to acquire and fluently manipulate any evolved orthography and a ‘marked’ periphery of skills that are only applicable to subsets of orthographies. An inn&e bias to ‘search’ the orthography of one’s native language for word-, syllable-, phonemic- and phonetic-level structure in that order would lead to fewer blind alleys than the reverse order of search. Most current work on disorders of reading i:; best regarded as pertaining to the algorithmic sub-thecry. Explicitly distinguishing stlch a level serves to remind m that formal algorithms must be constructed that are in principle capable of performing the operations that ale tacitly assumed in our block diagrams. The level of mechanism; the internal structure of computational centres and codes, should provide an alternative locus fog the interpretation of the rapidly increasing number of doubledissociations that are being uncovered in detailed singlecase studies (Shallice, 1979). To invoke a new box or arrow for each stimulusdimension and each input or output modality that can be selectively impaired just cannot be the way forward. And finally, we hope that good guesses at the level of mechanism will help us to interpret functionally the extremely exciting architectonic studies of language-committed cortex that have been appearing over the last decade.
2 14
J. C
Mar&all and E Newmnbe
Refen:nces &auvois, M. F. and D&ouesni, J. (1979) Phonological alexia: Three dissociations. J Neur@l.,!Veuroaurg., psych., 82, 1115-1124. Braak, H. (1980) Addtectonics of the Human Telencephulic Correx. Berlin, Springer. Coftheart, M., Patterson, K. E.. and Marshall, J. C. (eds.) (1980) Deep Dyslexia. London, Routledge and Kegan Paul. Fonder, K. I. (1978) Acceskg the mental lexicon. In E. Walker (ed.), Explorations jn the Biology of L.anguuge. Montgomery, Vermont, Bradford Books. Galaburda, A. M. and Sanides, F. (1980) Cytoarchitectonic organization of the human auditory cortex. J. compar. Neural... 190. 597-610. Geschwind, N. (1962) The anatomy of acquired disorders of reading. In J. Money (ed.), Reading Disorders. &&more, Johns Hopkins. Goldtimond, I. and Hawkins, W. F. (1958) Vexiervsrsuch: The logarithmic relationship between wordfrequency and recognition obtained in the absence of stimulus words. J. exper. Psychol.. 56. 457-463. Greenblatt, S. II. (1976) Subangular alexia without agraphia or hemianopsia. Br. Long., 3, 229-245. Klima, E, S. (1972) How alphabets might reflect language. In J. F. Kavanagh and I. G. Mattingly (eds.), Lungrrage by Ear and by Eye. Cambridge, Mass., MIT Press. Lichtheim, I,. (1885) On ap’hasia.Bruin, 7,433-484. Marr, D. (1’980) Visual information processing: The structure and creation of visual representations. PhZ Trans. R. Sot. London B, 290. 193-218. NlarshaIl, 5. C. (1976b Neuropsychological aspects of orthographic representation. In R. J. Wales and E. Walker (e&s.), New Approaches to Lunguage Mechanisms. Amsterdam, North-Holland PubIishing Co. .MarshaU,J. 1C.and Newcombe, F. (1966) Syntactic and semantic errors in paralexia. Neuropsychol., 4, 169-176. ‘Marshall,J. C. and Newcombe, F. (1973)Pattems of paralexia: A psycholinguistic appr0ach.X Psycho&g. Res. 2, 175-199. Morton, J. (1979) Word recognition. In J. Morton and J. C. Marsha8 (eds.), l?sycho,liflgustics Series, Vol. Z!.London, Elek. Moutier, F. (1908) L’aghosie de Broca. Paris, SteinheiI. Newcombe, F. and *Marshall,J. C. (In press) On psycholinguistic classifications of the acquired dyslexias,. Bull. Orton Sot. Sasanuma, 8. (1980) Acquired dyslexia in Japanese: Cliiical features and underlying mechanisms. In M. GDltheart, K. Patterson, and I. C. Marshall (eds.), Deep Dyslexia. London, Routledge and Kegan Paul. Schwartz, M. F., Saffran, E. M., and Marin, 0. S. M. (1980) Fractionating the reading process in dementia: evidence for wordkpecific print-tosound associations. In M. Coltheart, K. Patterson, and J. C. Marshall (eds.), Deep Dyslexia. London, Routledge and Kegan Paul. ShaiIice, T. (1979): Case study approach in neuropsychological research. J. clin. Neuropsychol.. I, 183-211. Taft, M. (1979) Lexical access via an orthographic code: The basic orthographic syllable structure (BOBS).J. verb. &urn. verb. Behav.. 18, 21-39. Teuber, N. L. (1955) Physiological psychology. An. Rev. Aychol., 9, 267-296. wass@ton, E. K. and Shallice, T. (1979) Semantic access dyslexia. Bruin, 202,43-63. Wanington, E. K. and Shallice, T. (1980) Word-fonm dyslexia. B&n, 203, 99-l 12.
Cog&& 1G (1981) 215-225 @ Elsevie: Sequoia LA., Lausanne - Printed in The Netherlands
215
Trends and debates in cognitive psychology GEORGE A. MILLER+ Prince ton Uniwrsity
In the past quarter century computer technology has led to developments that have transformed cognitive psychology. Nearly everyone has heard something about it, but not everyone understands what has been going on. One way to describe it is to say that the machines are catching up. When Satchel Paige was asked how he managed to last so long as a great baseball player, he is reported to have said, “Never look back-somethir+Jle may be gaining on you”. I often think of him in connectio:n with artifjcial intelligence, or “A..I.“, as it is known to people who love it. There are those who object that it di.minishes human dignity to compare people to machines. I think the truth is that the comparison now signals a new conception of what machines can be and 40. And what is that new conception? Stated abstractly, the modern computer has led to the concept of a physical symbol system (Newell, 1980)-“the concept of a broad class of systems capable of having and manipulating symbols, yet realizable in the physical universe” (Newell, 1980, p. 135). A computer is a physical symbol system. The basic assumption of A. I. is that a physical symbol system is capable of intelligent behavior. That is to say, the ability to accept input symbols and generate output symbols, to store and erase them, to compare them and to branch according to the outcome of the comparison are the only kinds of building blocks required for the synthesis of intelligence. The claim that physical symbol systems can be intelligent seems to have been well established by many examples. However, the further claim that human intelligence can be modelled by physical symbol systems takes this basic assumption an important extra step. And some would go still further to claim that the human brain ‘isnothing but a physical symbol system. Thl: sttempt to explore these further claims has led to the use of computers to sirnulate the cognitive processes of human beings, and has resulted in inany st udies that compare human and artificial intelligence. Most psychologists go abcsut their business, studying the thought and behavior of living organisms, without worrying about advances in computer the ory or technology. But those of us interested in cognition have been looking *I would like to express my appreciatiorl to Andrew Ortony, for his strenuous objections to this paper. Reprint requests should be sent to Dr George Miller, Psychology Department, Princeton University, Princeton, New Jersey 08544, U.S.A.
216
G. A. Miller
over our shoulders frequently in recent years. Indeed, we are fascinated by the possibility that we can understand human intelligence better by understanding better the structure and workings of intelligent machines. A. I. offers us an exact language in which to formulate theories of cognitive processes, along with a method for determining precisely the behavior that those theories will explain. The parallels between artificial and natural intelligence have enormously enriched contemporary cognitive psychology. It is not surprising, therefore, that many cognitrvepsychologists have accepted the claim that all human thinking is information processing and that many theoretical ideas can be transferred more or less directly from A. 1. to the description of human intelligence. This assumption has characterized one of the most productive lines of research in cognitive psychology in recent years. However, the idea that human intelligence can be modelled by physical symbol systems has not been accepted without criticism. Since it often helps clarify a discussion to consider objections to as well as arguments for a hypothesis, I want to discuss three of the many objections that have been raised. For convenience, I shall call them (1) the question of relevance, (2) the question of testability, and (3) the question of completeness. The question of relevance raises the possibility that artificial intelligence may be achieved in ways completely different from thorn that evolved with th : human brain. Testability leads into basic questions about the relation of lany!tage to cognition -can verbal reports of thought processes be used to t&4 the validity of compzlter simulations of those processes? What, if anything, do simulations of mental processes leave out is the question of completeness. I shall discuss these three questions in turn, and conclude with some comments on problems still outstanding. Relevance to human intelligence The question of relevance is nicely ihustrated by chess playing machines, which have been intensively and competitively studied by workers in A. I., probably because the performances of men and machines are so easiIy compared. As of this writing, the best machines cannot beat intrnlational grand masters, but they are sneaking up into the master class. The point I want to make here is that the programs that simulate the thought processes of human players have proved inferior to programs that simply exploit to the limit the sheer speed and power of modern computers. .From studies of chess *players (De Groot, 1965) it is known that on any given move in the middle game a grand master will consider only a few hun-
Trends and debates in cognitive psycho&y
2 17
dred alternative lines of play, and some lines may be explored in considerable depth. The most successful computer programs, on the other hand:, explore every legal continuation for about three moves ahead-millions of alternative lines, but none in depth. Improvement in the computer’s skill has resulted from larger and faster machines, not from cleverer heuristics for solving chess problems. The moral it that, if your goal is to use a computer to perform some function as mtelligently as possible, the best solution may not be to imitate expert human beings. In that case, an A. I. solution may be as irrelevant to cognitive psychology as the wheel is to an analysis of walking. Artificial intelligence may travel the same road taken earlier by formal logic (Domotor, 1978). As long as logic was thought to study laws of human reasoning, logicians were expected to explain logical fallacies as well as logical truths. Not until Frege cut this connection was formal logic able to develop, to become one of t:cle outstanding intellectual accomplishments of the 20th century. When A.I. similarly renounces psychologism, it too may take off toward accomplishments we cannot now even imagine. Those who believe that A.I. is relevant to the study of human cognition usually reply to this possibility by pointing out that chess programs are a special case, and that so far most advances in A.I. have been achieved by modeling them on what we know or believe about human performance. Of course, this situation could change. But if A.I. does develop non-anthropocentric kinds of intelligence, the general principles and specific devices that it will generate will still be useful in understanding human intelligence (Pylyshyn, 1978). Indeed, a truly general theory of intelligence will not be achieved until we can characterize human intelligence as a special case. The outcome of this debate, therefore, is that the goals of artificial intelligence and of cognitive simulation .are indeed different, and may become more so. But these differences raise no serious challenge to the central claim that human intelligence can be modelled by physical symbd systems. Testability: verbal reports as data The conclusion to the fast question indicates the importance of the second question. How is the special case of human intelligence to be recognized? If there are many ways to perform any particular intelligent function, how are we to know which way people perform it? The obvious answeiris to ask them, but thst turns out not to be as simple as it sounds. A history of scientific psychology could be written in terms of the debate over the use of verbal reports of mental processes. More than a century ago Wilhelm Wundt defined psychology as the science of immediate experience and proposed to explore it by collecting verbal reports of observers trained
218
G. A. Miller
in his special technique of introspection. But it soon became apparent that much of what goes on during perceiving, remembering, and thinking is simply not accessible to consciousness and therefore not reportable (Humphrey, 1951). When American behaviorists redefined psychology as the science of behavior, the unreliability of introspective reports Yrasone of the reasons they cited for abandoning consciousness as a scientific problem. Karl Lasnley made the case in a famous paper in 1923, vlhich he revised in 1958 and summarized as follows : “hb activity of mind is ever conscious, This soundslike a paradox, but it is none the les true. There are order and arrangement,but there is no experienceof the creation of iIut order. I could givenumberlessexamples,for there is no exceptionto the rule. A couple of illustrations should suffice. Look at a complicatedseenme. It consistsof a Tsumberof objects standingout againstan indistinct background;desk, chairs, faces. Each consists of a number of lesser sensations combined in the object, but there is no experience of putting them together. The objects are immediately present. When we think in words, the thoughts come in grammatical form with subject, verb, object, and modifying clauses falling into place without our having the slightest perception of how the sectence structure is produced ... Experience clearly gives no clue as to the rnea~~sby which it is organ%zed”. LasMey’s citim has since bee?1 echoed by several cognitive psychologists (Man-
dler, 197&z; Miller, 1962; tieisser, 1967); some have collected evidence in support of it _ NXxtt and Wilson (1977) reported sever31 situations in which people seemed to be unaware-of stimulus factors that determined their responses. For example, passers-by in commercial establishments were asked to evaluate an array of consumer goods. They were asked which item was the best quality and, when they announced a choice, were asked why they had chosen the article they had. The choices showed a strong position effect, such that the right-m%.t item in the array was preferred to the left-most item by a ratio of almost four to one, but nobody ever mentioned position as the reason for the cho.&. Indeed, when asked directly about position, everyone denied that it had had any influence on their decision. In this and several other situations Nisbett and Wilson concluded that only the pro&d of the mental process was accessible to consciousness; the process whereby the choice occurred was not open to introspective report. If this conclusion is correct, verbal reports about how people solve problems and make decisions canno! be used as data to test the psychological plausibility of artifically intelligent systems. Critics of this conclusion have replied, however, that the distinction between mental processes and mental products has never been clearly drawn
Rends and debates in mgnitile ps_vchulog~ 2 19
(Smith and Miller, 1978). As White ( 1980) remarks, it is all too is easy to fall into the trap of calling anything that appears in consciousness “product” and everything else “process”, in which case Iashley’s claim becomes true by definition. Smith and Miller (1978) would drtip the question of whether people have access to processes and focus insteac’ on when such access can occur. Introspections are supposed to be verbal reports of cognitive events. If the two are not closely related+t the reports are incomplete, inaccurate, or irrelevant-the prospects for cognitive psychology seem rather dim. Indeed, the whole effort to formulate psychological theories in terms of physical symbol systems would seem to be threatened, since there would be no way such theories could be tested. Surely, not all verbal reports of mental events are worthless. No doubt many verbal reports should not be taken as complete and veridical d12scriptions of me:ntal processes, but it seems absurd to claim that intelligent adults cannot say anything true or informative about their own conscious experience. The question is, when can we trust a verbal report? What factors determine the veridicality and usefu?ness of verbal reports? Such questions inspired Ericcson and Simon (1980) to undertake a clas sification of the different kinds of verbal reports. Reports of the outcome of some decision process will normally be accepted as veridical. If you ask me whether I would like an apple and I say “Yes”, you will accept my reply as veridical. To reject this kind of verbal report would be to deny the value of language as a medium for effective social interaction. Reports of events leading to such a decison, however, are harder to evaluate. Sometimes they appear to be total fabrications, as when people try to explain why they have conformed to a post-hypnotic suggestion. Others seem veridical, as when people report concurrently on the subgoaIs they are considering in the course of solving some complicated problem. Some factors affecting the value of verbal reports are obvious. Concurrent reports are more likely to be accurate than are retrospective reports. Reqorts about specific mental events are less likely to be fabricated than are comments that require the thinker to draw abstract conclusions about goals or methods. Based on their review of the literature, Ericcson and Simon concluded that only information in focal attention can be verbalized, and that information appears in focal attention only when information processing is executed under cognitive control, not when it is executed automatically (Shiffrin and Schneider, 1977). Ericcson and Simon list perceptual recognition, retrieval from memory, and skilled motor actions as examples of automatic processes. Information processing that has not become automatized, however, is accessible for verbaI report. Since most steps in solving unfamiliar problems will not be auto.matized, a concurrent verbal report can contain much useful data about
220
G. A. Miller
the processes that the human thinker is executing and the order in which they iMXW. The
outcome of this discussion of the second question, therefore, is that even though some kinds of verbal reports cannot be trusted, and even though some automatic mental prscesses are unavailable for verbal report, under optimal conditions it is possible to test the psychological plausibility of intelligent programs. That is to say, under optima! conditlJns it is safe: to use verbal reports as data. So the claim that human intelligence can be modelled by physical symbol systems need not be abandonned as untestable. Between Lashley’s original claim and Ericcson and Simon’s answer, however, something seems to have slipped. Lashley made a claim about the contents of conscious experience. Ericcson and Simon do not dicsuss conscious experience. Instead, they offer a model of verbalizing. They spea.k of the information that finds its way into short term memory as being heeded or attended to, and say that only the information in focal attention can be verMized. A natural assurnptior~ would be that they believe the contents of’ short term memory and the contents of consciousness are identical, but they are careful not to make tltat &im. A computer can simulate short tenm memory, but it is not clear how a computer would simulate consciousness. The compLetenbessof computational theories If a computer were used to model the weather, no one would fear that a cyclone might destroy the computing center. But using a computer to model cognitive processes is frequently assumed to be different: the brain is itself a computer in a sense in which weather is n,nt_
Trendsand debatesin cognitivepsychok:gy 22 1
computational theorists for having done no better. I assume that it will be much more difficuh to instantiate consciousness than to understand atttibutions of consciousness to others, but for the moment I will ignore the fact that we lack an adequate theory of either instantiation or attribution. The question for the moment is whether cognitive simulation could contribute to such a theory, or whether it must fail on principle to accommodate the conscious-unconscious distinction. It would not be surprising if we were to find that a cognitive theory was incomplete. Immanuei Kant popularized the famikar threefold categorization of mental faculties: coignitive, affective,and conative; or knowing, feeling, and willing. The implication is that a theory of cognition could not be a complete theory of the mind, since emotional and intentional dimensions would be excluded. But to ignore i\o,eling an8 willing is not the same as to ignore consciousness. What do we know about consciousness? A few fundamental propositions would probably inspire general agreement: Consciousnessoccurs in living systems Some would amend this to, “and only in living systems”--which the qusestion at issue here.
prejudges
Consciousness is associated with motility Presumably, rocks and trees are not conscious-which suggests that consciousness is important for volitional aspects of purposive behavior (Langfeld, 1927), and for the control of perception (Powers, 1973). The level of consciousness depends on activity in the iimbic system -On the general level of arousal, which suggests that consciousness is important for affective and emotional aspects of mental life (Mandler, 1975b). Consciousnessserves some useful purpose in cognition The argument here usually proceeds as follows: Consciousneu would not have evolved if it had1not had survival value; presumably, consciousness facilitated some way that cognitive processing-- made its possessor mord intelligent-in we do not yet fully understand, and intelligence had survival value. But perhaps consciousness is less important for cognition than we had formerly imagined. Ef it were true that a physical symbol system, for which the conscious-unconscious distinction is not relevant, could perform all the cog-
222
G. Al.Miller
nitive processes of an intelligent person, would that not imply that consciousness is unnecessary for cognition‘? Or, to word it more carefully, can one say that to simulate a cognitive function successfully is to prove that consciousness is unnecessary for the performance of that function? In my opinion, that would be a research program of considerable psychological importance. The very fact that complex cognitive processes can be executed unconsciously, assuming it is a fact, should provide an important datum for anyone interested in consciousness. How might psychological theory accomodate such a generalization? Perhaps ccnscioutness emerged in the evolution of affective and volitional systems, and provides no more of a window on cognition than is reqrlired for feelings and for purposive movement. Or-as seems more plausible to me-perhaps A.I. will prove to be an incomplete theory of cognition, a theory of certain lower level processing operatiorrr that require conscious attention only when they fail. I shall not pursue further the question of completeness, except to say that I cannot see how it can be settled until we have a theory of brain function adequate to suggest what else, other than information processing, a brain might do. And I need hardly remind anyone that we are fir from having such a theory of brain function at the present time. Presentation and representation The crowning intellectucal
accomplishment of the brain is the real world. Physicists and chemists long ago demonstrated that the real world of our experience is very different from the inanimate universe of physics and chemistry. The sounds and colors we perceive, the apparent objects that integrate them, the space in which those objects are located, the values we attach to them, the intentions we attribute to others -all these fundamental aspects of the real world of’our experience are adaptive interpretations of the really real world of physical sci.ence. I would like to use the word ‘“presentation” to refer to the way the real world presents itself to u.s or, more precisely, to the awareness we have at any moment of thig real world we have constructed (Ward, 19 19). All organisms achieve some presentation of their environments adequate for their survival as a species, although they do it in very different ways: the human world is very different from the world of a honey bee, but similar to the world of a chimpanzee. Moreover, categorization is a basic process in the construction of any such representation: at the very least, substances must be categorized as edible or inedible and organisms must be categorized as friend or foe. Insofar as we can discover something about the categories recognized by a spe-
Trends and debates in cognitive psy~holo~ 225
ties, we can come to appreciate something of the world in which it lives. The work requires great imagination, but I see no reason to conclude that it is impossible. For human beings, the presentation problem arises a1.t two levels: first, at the level we cab. the real world ; second, at the level of communicable symbols. I find it convenient to distinguish between percepltual presentations and symbolic representations. The level of symbolic representation not only builds on the cognitive categories established at the level of perlceptual presentation, but introduces many conventional categories that our ancestors have found useful. If the level of symbolic representation could be lxgarded as a simple one-to-one mapping onto presentations at the perceptual level, cognitive theory would be much simpler than it is. But the two levels interact so intimately and pervasively that i; may be misleading to try to pull them apart. I take the problem of characterizing the interactions between these two levels-between the real world and the world of words-to be the central problem in the study of human cognition (Mille:r and J,ohnson-Laird, 1976). Cognitive psychologists recognize that the symbolic representation influences the perceptual presentation in subtle ways--it influences what a person pays attention to and what perceptual distinctions will be drawn end remembered. Most important, it brings to bear cognitive schemata that enrich the perception and the person’s response to it. The symbolic component does not simply label the output of the perceptual analgzer; it also controls the input to it. Minsky ( 197 5) has proposed frame theory as a possible answer to such problems. A frame is a list of attributes associated with a colncept, along with default values for many of those attributes, Recognition is achieved by matching the input to the appropriate frame,; where perceptual input is lacking it can be supplemented by adopting the default value-the most likely value that instances of that category have had in the past. Cognitive theory will need schemata, or frames, just as certainly as it will IJ ed categories. And it may need several further insights before it can capture the intentional quality that is so much a part of our mental life. The matte; of what a person intends to do can be fruitfully discussed in computational terms (Miller, Galanter, and Pribram, 1960; Powers, 1973; Rosenblueth, Wiener, and Bigelow, 1943). Those of us interested in +he relation of language to cognition are constantly reminded of the importance of such intentions; people sb ,seldorn say what they really mean that if hearers could not attribute appropriate communicative intentions1 to them, linguistic communication as we know it would be impossible (&hank and Abelson, 1977). But the more pervasive intentional phenomenon is has~zally se:mantic: how do processes ir a brain intend (become symbolic of) something beyond themselves?
224
G. A. Miiler
In general, the adequacy of any computer simulation is related directly to the adequacy of our knowledge of the cognitive process that the system is intended to simulate. For example, linguistic understanding of phonology and grammar is reasonably advanced, and there we have systems of considerable power. Psychological understanding of communicative intentions, on the other hand, is still weak and groping, and there we have only systems of limited SCOpe. I suspect we will not make rapid progress in these more complex systems until we have solvtid two basic problems: fust, how to organize very large data bases the way people do, and, second, how to characterize the human point of view. The fti of these problems has been well enough defined so that we can expect to see considerable progress in the next few years. But the second is a vague and elusive ambition at the present time. To conclude: cognitive psychology has profited enormously from its interactions with the new data processing technology. Forty years ago psychologists interested in the so-called higher mental processes had few conceptual tools to work with beyond perceptual thresholds and chains of conditioned reflexes. Today we talk seriously about the organization of huge memories agd the overall structure of intelligent systems, topics that would have sounded like pure moonshine before they were objectively instantiated by the new technology. When I fwl discouraged about all the difficult problems that remain outstanding, therefore, I comfort myself with the thought of how far we have already come. The test of our progress is not the extent of our ignorance, but the extent to which we have accumulated knowledge that we can act on without fear. !f I extrapolate our progress at the same rate I have seen in my own lifetime, the future looks very bright indeed.
Refi?fella?s De Groat, A. D. (1965) &wghrt a& choie in chess. The Hague, Mouton. Domotor, 2. (1978) Al: model-theoretic aspects. Eehav. Br. Sci, I, 104,-105. Ericsson, K. A., and Simon, H. A. (1980) Verbal reports as data. Psychof. Rev., 87, 215-2511. Humphrey, G. (1951) %%ing: An inttorlucrionto its experimental psychology. London, Metheun. LangfebI,H. S. (1927) Consciousness and motor response. Psychol. Rev., %,I-9. L&by, K. S. (1923) The behavioristic interpretation of consciousness. Psychd Rev., 30, 237-272, 329-353. btsldey, K. S. (1958) Cerebral organization and behavior. In The bruin and human behavior, Proce.
e&g& of theAao&tbn for Research in lr!ervousand MentalD&orders,36,1-18. (Reprinted in F. A. Beach, D. 0. Hebb, C. T. Morgan, and H. W. N&en (edr.), The neur0p:~~~chdog.y of Lash&y: SeiiWedpapers~ofK. S. Lushky. New York, McGraw-W, 1960).
Trendsand %.-Sates irrcognifiw psychobgy
225
Mandler, C. (1975a)Consciousness: Respectable, useful, and probably necessary. In R. Solso (ed.), Information processing and cognition: The Loyola symposium. Hillsdale, NJ, ErEbaum. Mandbr, G. (1975b) Mindand Emotion. New York, Wiley. Miller, G. A. (1962) Psychology: The science of mental life, New York, Harper & Row. Miller, G. A., Gaianter, E., and Pribram, K. H. (1960) Plansard the StpLctureof Beh. vior. New York, Holt, Rinehart, and Winston. Miller, G. A., and Johnson-L.&d, P. N. (1976) Language and Perception. Cambridge, Mass., Harvard University Press. Minsky, M. L. (1975) A. framework forrepresenting knowledge. In P. H. Winston (ed.), The PsychoZogy of computer vision, New York, McGraw-Hill. Neisser, U. (1967) Ccgnitive psychology- New York, Appleton-Century-Crofts. Newell, A. (1980) Physicar symbol systems. Cog. Sci., 4,135-183. Nisbett, R. E., and Wilson, T. D. (1977) Telling more than we can know: Verbal reports on mental processes. Psychoi. Rev., 84, 23 l-259. Powers, W. T. (1973) Behavior The Control of Perception. Chicago, Aldine. ’ Pylyshyn, 2. (1978) Computat:onal models and empirical constraints. Behav. Br. Sci., I, 93-l 27. Rosenblueth, A., Wiener, N., and Bigelow, J. (1943) Behavior, purpose, and teleology. Philas. Sci. IO, 18-24. Schank, R. C., and Abelson, R. P. (1977) Scripts, Plans, Goa& and Understanding. Hillsdale, NJ, Erlbaum. Shiffrin, R. M., and Schneid x, W. (1977) Controlled and automatic human information processing: 11. Perceptual learn%, automatic attending, and a general theory. Psychol Rev., 84, 127-190. Smith, E. R., and Miller, F. D. (1978) Limits on perception of cognitive processes: A reply to Nisbett and Wilson. Psychol. Rev., 8.5, 355-362. Ward, J. (1919) PsychologikalPrinciples.Cambridge, Cambridge University Press. White, P. (1980) Limitations on verbal reports of internal events: A refutation of Nisbett and Wilson andof Bem.Psychol. Rev., 87.105-112.
Cognition, 10 (1981) 227-234 @Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
227
Will Cognitiob, survive? JOHN MORTON+ (MRC Applied Psychology l/nit, Cambridge)
The most important problem for cognitive psychologists is that of facing the complexity of the task we are engaged on. The key to this lies in the way we develop theories rather than in our experimental methodology. Not that the latter should be ignored, but without better notions of the mechanisms of the mind we cannot begin to understand what our data mean. Information procesdng models
An early decision for the theorist is that of deciding on a for-m of expression. My own decision in this respect has been to adopt the conventions of information processing modelling. The main principle here is to try to isolate processes which can be regarded as functionally modular. These processes are symbolised by boxes with inputs and outputs specified, usually in terms of the form of the n,ode they carry. The extent to which sub-processes are diffe.rentiated will depend upon the data currently being considered. Equally, the extent to which the mechanism of a particular process is specified will also depend upon the data being considered. There is the assumption, however, that the nature of a particular process has no bearing on the form of the model as a whole. Let me take one example. On the basis of a number of experiments on long-term facilitation effects on the recognition of Lc;histoscopically presented words my co-workers and I have proposed the existence of a modality-specific input lexicon, also called the ‘visual input logogen system’ (Morton, 1979). This process is responsible for categorizing a string of letters in terms of its morphemic content. It has inputs from processes which analyse the visual input and form central processes which use the available context. Its output is in terms of the morphol.ogical constituents of the input, and this output is passed to processes responsible for syntactic and semantic analysis. The data indicates that this input lexicon is separate from an equivalent process which performs the same function for speech inputs. At its most elementary, the resulting model is as in Figure I. The data seem to require a configuration of this form. However, the way in *I am grateful to Debra Bekerian, Lindsay Evett and Pat Wright for attempting to restrain me. I regret they only partially succeeded. Reprint requests should be s&t to J. Morton, MRC Applied Psychology Unit, 15 Chaucer Road, CambridgeCB2 2EF, England.
228
John Morton :
Figure 1.
The fragment of a model necessary to account for certainfacilitationeffects in word recognition.
level at which the facilitation effects take place
modality-free analyses
which the two ms!>dalityspecific categorisation devices operate has no bearing on the model as shown. Thus, one or both could operate as sets of simple, independent, evidence-collecting devices as I originally proposed (Morton, 1969). Equally, however, the auditory system could well have one or more of the characteristics of the cohort model proposed by Marslen-Wilson and Welsh (1978). For example, it could well be the case that negative information is used to initibit inappropriate conditions. Figure 1 would be unaffected by the decision. Of course, the model !n Figure 1 is only a fragment of the model I currently work with. Pre’zisely, it is just that fragment which is required by the fzcilitation data. Ther: would be other ways of accounting for the same data, but the modelling is i:-onstrained by a variety of facts which enter into the definition of the func:l:ions shown (see Morton, 1968, 1979). The technique of long term facilitation of the recognition of impoverished stimuli serves now to define furthe:; the properties of the input lexicons. Thus, Murrell and Morton (1974) s3owed that prior presentation of SEEN facilitates the subsequent recogniton of SEES. Prior presentation of SEED, on the other hand, has no effect on SEES. Thus we can conclude that the input logogen system is based on the morpheme rather than on the word. An unpublished paper by Steve Kempley and myself shows the same to be true for auditory stimuli. In addition there is no transfer between irregularly related words. Thus, having hea?d the spoken ‘bring’ has nc effect on the subsequent recognition of the word ‘brought’ heard in noise. Other experiments by Osgood and Hoosain (18.74) in vision and Gipson (unpublished) for audition show ,
Will Cognition survive?
229
that stockmarket has a unitary representation at this level but street market does not. We should note here the need for careful replication not just in the conditions of stimulus presentation but also in the state of mind of the subjects. What we have to do is to guarantee that the subject:: are behaving in a controlled way, and are only using the resources which we think we are testing. Let us have a thought experiment. Suppose, one day, I read out to a subject the words ‘lion, tiger, zebra, hippo, rhino’ several times, telling the subject that these words will come in useful on the following day. Next day the subject is seated in front of a tachistoscope in the same room with the same experimenter. ‘See if you can guess the word I am showing you’, he is told. The first item is a short word beginning with H. ‘Hippo’, says the subject. Would we then want to say that the subject has shown long-term cross-modal facilitation effects in word recognition? And would this experiment constitute a non replication of experiments where the relation between the tasks was less clear to the subjects, where the subjects were instructed not ta guess and where no cross-modal effects were found? The answers could only be positive in the case that a superficial view were taken of our objectives. fn terms of the underlying reality --i.e. the hypothetical constructs of the model-the two experiments have little in common. They encourage different strategies in the subjects who, then, are calling upon different resources. A possible diagram of the processes involved in the second task is given in Figure 2. Missing from this is the control device w’%ch determines when and how the appropriate memory record is accessed and used; but this doesn’t present any conceptual difficulty. The task for the cognitive modeller is to express the two tasks in a common framework. This is sketched in Figure 3, where the central component is clearly in need of enormous expansion. Whst the model lacks, as shown, is a statement as to the conditions under which one or another route is used. What does seem to me clear is that the data can only be interpreted in terms of a model, that the model should gi;/e an account of a variety of tasks and that we should expect different tasks, or variants of the same task, to have different descriptions in terms of the model, The apparent chaos in the data on word recognition will only be resolved when we treat separately such tasks as lexical decision, word monitoring, reading time, semantic classification and t-scope recognition. To regard them all as equivalent with respect to the study of Word Recognition is a recipe for disaster. Our models should point plausibly to the differences between the tasks in addition to the similarities. Equally, a mcdel which accounts perfectly for, say, the data of lexical decision, but which cannot, in brinciple, account for, say, the interaction of context and stimutus information in recognition, must be of less value than one which can account for
23Q
John Morton
Figure 2 . The fragment of a model required to represent the use of a guessing strategy in responding to partiblly seen words.
memory
1
response
records
Figure 3.
A means of representing in the same model the effects represented in Figures I and2.
records
1
response
both, even though the Mter may be more complex. There will be an indefiiely large number of ways of representing 3 particular set of data. Only by extensive use ,f converging operations can we co&rain our choice in a reasonable fashion. It follows from what I have said that the real objects of study for a cognitive psychologist are the cognitive functions. The experimental techniques
Will Cognition survive I’
23 1
and the data which thiese yield must simply serve to illuminate the processes and not become objects of study in their own right. The same principle applies to dependent variables. Thus when we find that the variable of word frequency applies to recognition, reading time, lexical decision and picture naming, we can 2~ longer assume that the effects in all these tasks are to be located at the same point in our model. Word frequency is going to be reflected in the frequency of experience of a word (and so, possibly, in any recognition process) in the frequency of production of the word (and so, perhaps, in the ease of mobilising or producing the spoken response) and in the diversity of its associations (and so in any task sensitive to this). Attempts to treat all frequency effects as equivalent seem perverse given the complexity of the models we already have.
Alternative models Information paocessirig models are probably only useful at a particular level of theorizing. They seem particularly powerful for relating together a wide variety of data. Sometimes, for a more detailed study of a particular task (if one has reason to do this) ~5 for a more precise description as to how a particular function is impleilented, then other notations may be preferable. What is necessary is that the objectives be clearly defined. Thus we should distinguish clearly between an inft3:rmation processing model of a set of processes and a flow diagram of the successive operations of these processes. The latter is exemplified in the analyses of picture-sentence matching (Chase and Clark, 1972). These are legitimate analyses, of course, but unless the operations depicted in such flow diagrams are linked to functions which operate in other tasks, their utility seems limited. The second type of altenlative model involves the more detailed specification of the operation of a process. For this, one of a variety of notations may be used. Production systems and augmented transition networks (ATNs) are two such notations. What we must beware of here is believing that the notations themselves have model status. There may be model implications in any particular production system (for example* ir the u’sy the memoy{ buffer is handled), but productions and ATN’s are not themselves models. Neither is ‘psychologically real’ (Bresnan, 1978); they are just means of expression which may be more cr less convenient for particular circumstances. The result will 3e a model, but the model status of a simulation lies at a level above the specific implementation in the same way that any computer progre h.as its essential functional description which is independent of the language in which it is written.
232
John Morton
Levels of &eory
In the previous section I made a distinction between model and notation. The importance of this distinction resides in the differences in the nature of the scientific discourse which are relevant to the two. Basically, a notation cannot be falsified. This is as true for the information processing notation as for Productions and ATN’s. Models, on the other hand, can be falsified if they incorrectly represent the data. Thus a theory can be falsified without affecting the status of the framework (or notation) used in the model, The justification of the framework springs from its utility which has three components, communicability, generality (already discussed), and applicability. Applicabilityof thstory
Experimental psycnology has a disastrous history with respect to its relevance. The lack of applicability of the verbal learning literature to real learning Situations is the most often quoted. Those of us who live by application of our work have long ago seen the advantages which accrue from looking at real tasks as well as those in the laboratory. To start with it prevents us from taking too narrow a view of human processing capabilities. Secondly it provides a sourtze of data wLich can feed into the models we create for laboratory tasks. 0ne of the successes of information processing models for word processing can be found in the way the models can be used to describe the effects of brain damage. Specifically, the models are being used to provide a new kind of taxonomy of the effects of brain damage which appears to have a productive and systematic power lacking in previous taxonomies (Patterson, 19s 1; Shallice, 1981). Furthermore, a number of theorists use data from brain damaged patients to justify and refine their models (Coltheart, Patterson and Marshall, 1980; Marshall and Newcombe, 1973; Seymour, 1979). A genuine extension of this trend (other than the typical addition to a grant proposal whereby any study of word recognition is represented as ‘having clear implications for the teaching of reading’) can only improve our theories of cognition”.
The danger
of “Cognitive Science’
Finally Hwould like to point to what I see as a trend which is in danger of becoming a cult. ‘Cognitive Science’ is an attempt to marry cognitive psychology, artificial intelligence and neuro-biology. There seems to me to be a
Will Cognition survive ?
23 3
right way and a wrong way to do this. The two dangers for cognitive psychologists are that they be led into one of two activities which can become el;ds in themselves. 1. Simulation 2. Reductionism It is clear that the theoretical repertoire of a cognitive theorist can be greatly increased by a knowledge of devices of all kinds. Compu,ters are particularly seductive in this respect. It could be argued that the computer unalogy had a very serious effect on psychological theory in thr: ‘60’s and ‘70’s by providing the strong analogy of the central processor. This led to focussing on single channel hypotheses and general purpose computation devices to the excessive exclusion of more distributed computation. This arose, in part, from the, then, impressive simulation of human-We activity on computers. A similar error could occur with the new generation of machines. Further, simulation could become an end in itself for psychologists. The argument seems to be that with tasks as complex, say, as language comprehensicn or object recognition, there can only be one solution as to how it is done. The computer solution is then going to correspond to the human solution. The flaw here lies in the definition of the problem. If the computer does not make th: same kinds of errors as the human, then the equation is difficult to justify. At best the claim :hat the simulation is humanoid can only be justified at the level of description of the model which is appropriate for the detail of data considered. And even this would only be the case if it could be established that all other possible models, at the level of concern or above, in principle could not accomplish the task. Inasmuch as thk has been attempted it has rather been on the failure principle. Thus we find that since no-one has succeeded in segmenting speech by computer by a bottom-up process (i.e., on the basis of only the speech signal without using syntactic, semantic or pragmatic constraints) the claim is made that the human system cannot do it this way. Now the failures of computer speech recognition clearly constrain us. Certain kinds of bottom up recognition devices are ruled out (specifically, those which failed) if one requires all of one stage of processing to be successfu!\y completed before the next begins. But there is nothing to prevent us fro:n supposing that the auditory input lexicon in Figure 3 is a purely bottom-up device which does the best job it can with further devices cleaning up the representation on the basrs of the various constraints. (There are actually good psychological reasons for believing this not to be the case, but these are irrelevant for the argument.) The constraint from A.I., then, on cognitive theory seems weak. The danger I see is that X.1. will become a substitute ;zct&Gtyfor people whose abilities are needed in mainstream cognition.
234
John Morton
The danger in the link with the neuro-sciences is more subtle. Basically it is that it will be thought that the only proper aim of a cogmtive theorist is to explain how his constructs are actually implemented in the brain. Mehler and Morton (in press) have discussed this issue at length. I will restrict myself here to the claim that the achievement of a purely psychological description of cognitive processes is an end in itself. The mapping of this description onto the brain is an equally worthwhile but different activity. To have the two activities confused would also take energy away from the purely psychological task. It is claimed that biological facts must constrain cognitive theory. I have not seen such claims substantiated. Condusion
There seems to me no need to doubt the future of cognitive psychology. Our models will become more complex and we must learn to relate together models of different kinds which aim at doing different jobs. We must avoid being diverted from the proper task and learn to develop in theoretical sophistication. This way Cognition will survive its editors. References Breman, J. (1978) A realistic transformational grammar. In HaIle, M., Bresnan, J. and MiBer,G. A. (eds.), Linguistic Theory and Aychologikal Reality. Cambridge, Mass.,MIT Press. Chase, W. G. and Clark, H. H. (1972) Mental operations in the comparison of sent:.nces and pictures. In Gregg, L. W. (ed.), Cognitionin Learningand Afemory. New York, Wiley. Coltheart, M., Patterson, K. and Marshall, J. C. (eds.) (1980) Deep Dyslexia. London, Routledge and Kegan Paul. Marsha&J. C and Newcombe, F. (1973) Patterns of paralexia: A psycholinguistic approach. J. Psycholing. Res., 2. 175-199. Ma&en-Wilson, W. D. and Welch, A. (1978) Processing interactions and lexical access during word recognition in continuous speech. Cog. Aychol., 20. 29-63. Meek, J. and Morton, J. On reducing language to biology: Playing with your language organ. Unpublished ms. Morton, J. (1968) Considerations of grammarand computation in language behavior. In Catford, J. C. (ed.), Studies in Language and Language Behaviour, C.R.LL.B. Progress Report No. VI. University of Michigan. Morton, II. (1969) The interaction of information in word recognition. Aychol. Rev., 76, 165-178. Morton, J. (1979) Word Recognition. In Morton, J. and Marshall,J. C. (eds.), PsycholonguisticsSeries I!. London, Elek Scientific Booh. Murrell,6;. A. and Morton, J. (1’974) Word recognition and morphemic structure. J. exper. Aychol., 102.963-968. Osgood, C E. and Hoosain, R (1974) salience of the word as a unit in the perception of language. Percep. kychophys., I&168-192. Pattemn, K. (1981) Nemopsychological approaches to the study of reading. Br. J. Psychol;; 72, 151174. f%‘mo& P. IL K. (1979) HLmm VisualCbgnition. London, Collier Macmillan. Shall@ T. (1981) Neurological impairment of cognitive processes. Br. Med. Bul., 37, 187-192.
lrl (a9sr) 235-233 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
Cognition:
235
The LNR approach to human information processing DONALD A. NORMAN* DAVID
E. RUMELHARJ
University of California, San Diego
Our goal is to understand the mechanisms of the human information processing system. We view the human as an info.rmation processing mechanism built upon a biological substrate- the result of :many years of evolution and change, existing within and influenced by a social and cultural environment. We want to understand the t:asic information processing mechanisms that pick up environmental information, that store, retrieve, reconstruct, infer, deduce, and otherwise process that information. The goal, moreover, is to create a precise theoretical formulation of psychological mechanisms and knowledge structures. This means that our enterprise is multi-faceted: simultaneously examining a broad set of issues from across the range of psychological phenomena. Three different issues are dominant (and inseparable). First is the understanding of the psychological mechanisms that underlie behavior; second is the representation of knowledge and the many ways is which it is used by the psychological mechanisms; and third is the need to relate the mechanisms and knowledge within the person to the environment, the social interactions, and the culture. Human behavior results from the continual interaction of all of these influences, as much shaped by the cultural history of the person and the immediate demands of the environment as by the limitations and powers of the psychological mechanisms.
Levels of analysis: observations, fmntworks, formal models In moving toward explicit scientific theories, we operate at several; different levels, from observations, to the development of frameworks for research, to specific, detailed mathematical and simulation models. As we approach a new area, we first step back and attempt to identify the critical phenomena. Then, we develop a high level, theoretical framework to guide our work, to lay out appropriate directions of research, and to be suggestive about the end product. -*Reprint requests should be sent to D. A. Nomtan, Programin ?ognitive Sckiice, University of California, San Diego, La Jolla,Califoniia 92093, U. S. A.
235
Donald A. ilkmnan and David E. Rumelhart
Identifying the criticalphenomena One fruitful approach is to be observant of real phenomena, to ask questions of ourselves and of others, and to questi.on the psychological basis for even the most innocent observations. Thus, some of our work has been triggered by simple questions. ‘What was Beethoven’s telephone number?’ This is typical of the initial questions we ask. How is it that people know SO rapidly that they do not know the answer to such a question? This simple observation leads to some non-trivial results about the nature of knowledge, meta-knowledge, and retrieval strategies. ‘Where is the Empire State building?’ This question, fast posed to us by our students Marc Eisenstadt and Yaakov Kareev, poses another issue, this time in speech act theory and sociolinguistics: to answer the question properly requires knowledge of why the questioner wants to know and how much knowledge the questioner has. One of us speculated on possible routes to phrasing the answer properly, and the other developed a ‘room theory’ to provide a first order approximation to the proper answer. ‘What could it possibly mean to have a ‘pointer’ in human memory?’ This question led to the development of memory reference schemes and the notion of desi:riptions, issues that get at the way in which memory is structured.’ ‘Why did I do that? What caused me to make that error?’ This led us to observe the errors made by people in their everyday activities which in turn led us to study the relationships among action specification, performance, conscious and subconscious control mechanisms, memory representations, and memory retrieval. Developing general frameworks These observations, and others of a similar nature led. us toward the devk opment of framteworks,overall structures for viewing particular scientific puzzles. These are suggestive of the form of theory that will be required and of the data that must be obtained. These frameworks are of critical importance in forcing us to ask the right questions, to focus on the right areas, and to de%is
work was done by Bobrow and Norman.
The LNR approach to human information processing
237
veiop the right models. Here are some of the areas for which we have developed frameworks: attention and will; understanding; reading; perception; resource limitations; memory retrieval; learning; comprehension. Developing formal models Next, we delved into the topic area, attempting to specify the models as explicitlv as possible. Our goal here is complementary to that of the framework. Within a framework, we wish to be as broad as possible. Within a model, we wish to be as explicit as possible; and if this requires some overall simplifications or the exclusion of certain phenomena, for the moment, that is acceptable. The goal of model building is, in part, to provide an existence proof that the framework is viable, knd that explicit detailed mode!s of a particular area of cognition are possible within the framework. Within the model, however, we still wish to be broad. We fiercely resist the temptation to make different models for each different phenomena. We believe the mind works according to common, general principles; and although we argue that people have specific, speci&ed knowledge of the different areas in which they perform, we believe that the basic principles of the psychological mechanisms that underlie the specialized knowledge structures must be the same. Thus, the general theoretical structure for our model must apply to as many areas as possible. This sometimes means that when we leave one area of research and start the investigation of others, the developments within the newer areas may require us to reconsider the results from the earlier ones, to go hack over previous work and rethink the conceptual, theoretical bases. Our approach, therefore, is first to determine the critical phenomena, theti to specify the framework of th,: approach, and then to lay out the appropriate theoretical and experimental issues that will be examined. Theory is developed as appropriate, sometimes before the experiments-thus guiding the design and the experimental issues-ti sometimes afterwards, thus being guided by the results, and sometimes relatively independently, as when experimentd results, have as yet no interpretation and theories as yet no empirical tests. The OCcasional decoupling of theory and data are not of concern, because in the long run, the match will be made. Two themes: knowledge structures and processing mechanisms Two dominant issues recur throughout our studies. One concerns the nature of knowledge structures. The other concerns the nature of the ccWo1 and processing mechanisms used by the human information processing system.
238
Donald A. Norman and David E. Rumelhart
The human is a complex being, often simultaneously pursuing many different goals and act.:ionsequences. The environment provides a rich source of sensory information. Our movements are mediated through an elaborate and flexible set of muscles and tendons. In seeking to determine how the human can deal with the richness of both input and output, both themes come into play. The knowledge ;tiqlctures are of critical importance for all human activities, because they determine the contents of our perceptions, the structure of our thoughts, and the kinds and forms of actions that we can perform. Perhaps the major thrust within our laboratory revolves around the concept of schema, attempting to generalize its usefulness to cover the entire spectrum of psychological issues while simultaneously specifying its nature in a precise set of models for different domains of interest. We see the schema as an active processing data structure that organizes memory and guides perception, performance, and thought. The status of the schema today is halfway between being a framework and a formal, quantitative theory. In some of our work, the term has been used as a general organizing concept -primarily as a guiding framework for further analysis. In other work, we have spelled out in detail some of the organizational and processing properties necessary for the domain under study. The second major theme of our work concerns the problemsof the control of processing. Clearly, more data enter the human sensory organs than can be assimilated at any one time. Clearly, more topics and actions are potentially available for thought and action than can possibly be done at any one time. Some sort of selection mechanism is required: hence our interest in the problems of selective attention. Attention is but one aspect of the control of processing. Just how the various sourcesof knowledge and data interact has been a major theme of our research for the past five years, leading first to examination of data-driven (bottom-up) and conceptually-driven (topdown) directions of control, but lately more to the study of how different sources of knowledge can interact in constructive ways. Our work has progressed, first to provide a framework that demonstrated how schemas could examine arriving data and select among competing hypotheses, then to develop a detailed model of the interactive process that provides explicit mathematical predictions of a variety of phenomena in reading.? This work shows how partial information interacts with data structures and which, through inhibitory and excitatory links,, eventually converges upon unique interpretations of the percept. Moreover, this new approach is now seen to be part of a general class of interactive data structures that was explored to a great extent by Geoff Hinton while he was a postdoctoral fellow with ---__ 2Tti work has been done by McClelland and Rumelhart.
The LNR approach to human infimnah’ofl processing
238
us (the beginnings of this work are presented in the conference proceedings edited by Hinton and Anderson, 198 1). Perhaps most importantly, this work demonstrates that a system can appear to obey and follow general rules of language even though it does not have those rules within it. Thus, in the model of word perception, the system behaves as if it understood the orthographic rules of English, even though it does not. What the system does have, however, is a data base that consists of a large portion of the actual lexical items of the language, and the interactions of the processing mechanisms with the knowledge of the individual words makes the system appear to have general rules.
The future The field is entering a new phase, with Cognitive Science emerging as a new discipline, emphasizing the need for cross-disciplinary research, for a broader perspective on phenomena and greater tolerance by psychologistsof a variety of experimental and theoretical .methods. We expect to see new quantitative developments, especially as developments in computers move us away from simplistic notions of computation that no~v exist, towards richer possibilities, based in part upon our understanding of brain mec~nrnisms. Such systems will consist of simple computational methods to be applied in parallel, through thousands of processing structures. More depth is required in our understanding of representation, both to expand upon the work already performed and also to incorporate images,actions, motives, beliefs, and emotions. We must try to understand different levels of processing in order to encompass and understand conscious an4 sub-conscious phenomena and processes. In addition, we need to understand ;ne relationships of the human to the environmental influences and to cultural and biological origins, perhaps along the lines that we have suggested as a goal for Cogni+‘-re Science (Norman, 198 1). Our research group (the LNR research group) is now moving toward the study of new computational methods. We are enriching the notion of the schema by showing how different constraints can operate upon a system to yield an interpretation, without any explicit inferential or judgmental processes, Moreover, we now realize that there is a class of computational, cooperative techniques that show how cognitive processing can interact and settle upon a unique configuration, even when each of the basic sources of evidence is incomplete and ambiguous, for each source of evidence can be seen as offering rLconstraint upon the final interpretation, and the total constraints are sufficient to impose a unique interpretation. Thus, much of cognitive pro-
240
DorwidA. Normanand DavidE. Rumelharr
cessing may take place through simultaneous cooperation among many data structures rather than through the more traditional, serial inferential process now being considered. This idea is still little more than a topic of conversation with us, but it may guide our work in the coming years. Hints of this development can already by seen, however, i? the model of typing that we have developed, in the interactivation activation model of reading, and in the framework for the study of attention that we are developing. Finally, we expect more interaction of theory with applications, because the major problems of today deal with larger issues that require understanding of real behavior in the rich setting of the natural environment rather than the limited and restricted environment of the laboratory experiment. This will both enrich our understanding of psychological phenomena and also move the scientific pursuit of psychology closer to the development of useful tools and procedures for tackling real problems. References Bobrow. D. G., and Norman, D. A. (1975) Some principles of memory schemata. In D. G. Bobrow and A. M. Collins (eds.), Representatfon and Understanding: Studies in Cognitive Science, New York, Academic Press. Hinton, G., and Anderson, J. (eds.) (1981) ParallelModels ofAssociative Memory. H&dale, NJ., Erlbaum Associates. Norman, D. A. (ed.) (1981) Perspectives ori Cognitive Science: The La Jolla Conference. Norwood, NJ., Ablex Publishing Corporation; Hillsdale, NJ., Lawrence ErlbaumAssociates. Rumelhart, D. E., and McClelland, J. L. (In pre:;s) An interactive activation model of the effect of context in perception, Part 1. Psychol. Rev.
Cognition, 10 (1481) 231-242 @ Elsevier Sequoja %A., Lausanne - Printed in The Netherlands
241
Modularity as an issue for cognitive scietmce DANIEL
N. OSHcfGON
Massachusetts Institl
*
te of Technology
Several decades of research within disciplines now counted among the cognitive sciences has prepared a serious issue for investigation. The issue concerns the degree of modular structure exhibited by the system of faculties comprising the human intellectual repertoire. To clarify the modularity question it helps to construe faculties as putative kinds; so construed, faculties are maximal collections of psychologi,d. *_I1 p-recesses and structures that function in accord with a relatively small set of interlocking and explanatory principles. This view of faculties harbors two potential complexities that may be mentioned here and then put aside. First, it is commonplace to observe that mental process :s and structures may be manifest at several distinct Ievels of ‘reduction’, including the neurophysiological, the psychological, and the functional; so it is possible that processes or structur-zs grouped together at one level may fall info different faculties at another. In the latter ctise, it will be necessary to relativize one’s claims about faculties to particular reductive. levels.’ Second, it is logically possible for there to be no faculties at all in the desired sense. That is, the human nervolls system may offer up no coherent class of processes or structures susceptible to explanatory analysis by a relatively small set of principles. It has thus been a nontrivial achievement of the cognitive sciences to uncover a surprising degree of coherence in several areas of human ability, in particular, in language and vision; whatever prin-, ciples explain the regularities so far uncovered, it is plausible to assume that one or more faculties are implicated. These qualifications aside, the modularity question may be stated in gross form as follows; How many human faculties are there? An answer to this latter question will rest on answers to more refined questions concerning the principles governing various processes, structures, and functions. Faculties may be individuated on the basis of the incompatibility of the principles governing each. To clarify, let C1 and C2 be two classes of processes and *Reprintrequests should be sent to D. Qsherson, 2OC-124 (DSRE), MIT, Cambridge, Mass., 02139, U.S.A. ’ See Osherson and Wasow (1976) for discussion. Anothr ; potential complication: The acquisitional mechanisms for various faculties need nott, by logic, partition themselves in the same fashion as the acquired faculties, although one would susplect that a congruence obtains in fact.
242
DankfiX Osheron
structures that conform to two sets of interlocking and explanatory principles, P1 and P2, respectively. If the properties of C1 can be proved hot to be deducible from Pz, and likewise for C2 and P, , then distinct faculties are (provisionally) revealed. Note that to establish a claim for distinct faculties, it is not sufficient L- exhibit classes of processes or structures for which distinct explanatory principles exist; for, although the theories may be distinct, they might nonetheless be equivalent in the sense that it may be possible to reformulate one or both as a more general account of both kinds of phenomena. It is thus required .to provide a formal argument that processes and structures governed by one set of principles cannot be governed by the other. The next ten years of research may well see definite, if limited, progress in answering the modularity question. Explanatory principles are beginning to emerge in several subfields of linguistics, in visual pattern recognition (as in the work of Ullman (1979)), and in inductive logic (as in the work of Horwith (1981)). AS theories in these and other areas are articulated more and more precisely, relevant principles can be compared in the manner suggested above. No doubt initial comparisons will reveal a cloudy picture of partial overlap and independence among the several subsystems comprising putative faculties. With explicit claims about modularity in hand, however, we can hope for steady clarification. Of the several inquiries that might be undertaken by cognitive scientists over the next decade, investigation of the modularity issue seems to me to be among the most feasible and intrinsically interesting. References Horwich, Paul 11981) fiobubility and Evidence,Cambalclge. Cambridge University Press. @*herson, D-niel and Wasow, Thomas (1976) Species specificity and task specificity in the study of knguage: a methodological note. Cog., 4, 203-214. Ulhnan, Shimon (1979) TheInterpret&ionof VisualMotion. Cambridge, MIT Press.
Cognition, 10 (1931) 243-248 @ Elsevier Seqr:oia S.A., Lausanne - Printed in The Netherlands
243
What spatial representationand languageacquisition don? have in common STEVEN PINKER* Harvard University
For several years I have been studying two cognitive abilities: the mental representation of three-dimensiona. visual space, and the child’s acquisition of the syntax of her first language. When I describe these projects, the question I am asked most frequently is “What connections do you see between them?” I hope to use this note as an opportunity to answer that question as best I can. One answer to the question is that I have found it fruitful to pursue similar goals in the two research prog;ams. For example, the research projects are aimed at explaining cohesive cognitive capacities; I have littempted to construct explicit computational theories of those capacities; I have tried to use the specifics of the theories to raise empirical questions that have become the focus of experimental research programs; and I hope to evaluate the theories by seeing whether they provide re:!sonable accounts for as wide a set of findings relevant to’ the domain as l;oDsible. Rut a more interesting reply to the question would state wh.a~.-patial representation and language acquisition have in common at the level c.” actual cognitive mechanisms, not the way i[ have chosen to study them. A5er describing the re!;earch briefly, I will summarize my current conclusions about possible common mech.anisms in syntax acquisition and visuospatial cognition. My interest in spatial representation began when I discovered a paradox in the way people use mental images of three-dimensional objects and scenes. Mental images have been likened to pictures by laypersons, philosophers, and experimental psychologists alike, a view wh+o3emost plausible formulation r*n’p fl!V 7) computational theory of imcan be found in Kosslyn and Shwa,.-. 3 4. agery. According to that theory, im rgzP ;;re d%%utions of activated cells in a two-dimensional array which rece ~7s input from the eyes during perception and from long-term memory during Imagination. But if images are like two dimensional patterns, why are people so good at recreating the three-djmensional layout of a scene in their images, and at simulating smooth 3-D transformations such as rotation and translation (e.g., Shepard and Metzler, 1:97 1; Pinker, 198&z)? On the other hand, if images are more like three-dimensional c-ience Foundation grant BNS 80-24337. Re*Preparation of this paper was supported by National ._print requests should be sent to S. Pinker, Psychology and Social Relations, Williamhmes Ha& hmrd University, Cambridge, Mass., 02138, U. S. A.
244
S. Pinker
models, why does a mental image seem introspectively to be like a perspective glimps!: from a fixed vantage point, containing objtcts that appear to be foreshortened or occluded, or to be looming large as they approach (see Pinker and Finke, 1980; Pinker, 198Ob)? My experimental program has attempted to verify and extend these informal observations using chronometric and psychophysical measures developed by Shepard ( 1978)and his colleagues, Kosslyn (198!I), Posner (1980) and others. These studies attempt to assess how accurately people can detect both intrinsic and perspective-specific spatial relations among perceived and imagined objects before and after imagined transformations in the ‘third dimension (e.g., Pinker and Kosslyn, 1978; Pinker, 198Oa, in preparation; Pinker and Finke, 1980). We have found that people can do so quite accurately, at least for simple well-learned arrangements of objects. The theory I have been developing (Pinker, 198Oc, 198 1a) is inspired by the work of Attneave ( 1972, 1974), Kosslyn and Shwartz (1977), Marr and Nishihara (1978) Ullman (1979), and others. It posits that visual images and percepts are representations in a threedimensional array of cells, similar to Marr and Nishihara’s “21/2-Dsketch”, but with innuts both from the eyes and from long-term memory. Each cell represents a particular depth along a particular line of sight, and activating a cell corresponds to signaling the presence of a visible surface at that location. Each cell is addressable by its coordirlates in two coordinate systems: a fixed, viewer-centered spherical coordinate system; and a moveable, world-centered Cartesian coordinate system in which (unlike the viewer-centered system) a given range of coordinates corresponds to a constant real world extent regardless of which cells (those representing near or far surfaces) are involved. The dual addressing system is motivated by the insight that many visual processes involve translations between viewercentered and world- or object-centered coordinates (Hinton, 1979; Marr and Nishihara, 1978); in the dual-address array theory, this is accomplished by filling array ceils according to their addresses in one coordinate system (either vieweror world-centered), and then accessing the resulting pattern in the array using the other system. By appealing to this capability or::: can provide reasonably satisfactory accounts of certain aspects of top-down and of bottom-up pattern recognition, and of mental image generation, 3-D mental transformations, perceptual stability, visual attention, and the visual field-visual vc’orldalternation in perception (such as when we can see railroad tracks either as converging toward the horizon or parallel at every distance; see Gibson, 1950). Currently I am gathering experimental data in an attempt to test certain features of the theory and to flesh out its currently unspecified parts. Some sample questions: How, when we execute mental image transformations, do we fill in the newly visible details resulting from a rotation indepth,a “zooming
Spatical representation and language acquisition
245
in”, or a translation from the periphery into the center of the visual Geld? Which transformations can be used to shift the world-centered coorc!inate system relative to the fixed viewer-centered one, and which of these shifts of perceptual reference frame are under voluntary control, passively driven by the input properties, or both? Can visual attention be alloicated to any circumscribed reg$on of the 3-D visual space (corresponding to priming a cluster of adjacent cells in the array), or only to a cone centered on a particular line of sight (corresponding to priming a column of cells with silmilar viewer-centered addresses)? The goal of my work in language acquisition is a theory of 1he mental mechanisms children use to induce the syntactic rules of their language, basecl or the sentences they hear their parents use. This goal is different from, though complementary to, the goal set by most language acquisition researchers (e.g., Brown, 1973), which is to trace out the developmental sequence of language skills in preschool children. Instead, the most direct attack on this goal can be found in the work of linguistics and cogni:ive psychologists such as Wexler and Culicover (1980,), Anderson (1977), and the contributors to Baker and McCarthy (1981), and in earlier attempts reviewed in Pinker ( 1979~). One result seems to have emerged from all these attempts: a modeli’s success in accounting for language acquisition in a psychologically plausible way is in direct proportion to the number and specificity of the innate constraints built in tc the model.1 see this result as one tentative source of confirmation of Chomsky’s nativist proposals about language acquisition (e.g., Chomsky, 1962), tentative bacause it is possible (though, in my opinion, highly unlikely) that someone will devise a successful and plausible language learning algorithm lacking such innate constraints. Though the development of increasingly constrained theories of grammatical knowledge (e.g., Chomsky, 1979; Bresnan, 1981) has made the task of explaining language acquisition far easier, many difficulties remain in specifying how the putative innate rule schemas are modified in the requisite way by the linguistic data available to the child, One problem is that if current theories of grammar are correct, syntactic knowledge is couched in a mental vocabulary distinct from that underlying the perception of phonetic sequences and the perception of situations- the two types of potentially acquisitionrelevant input to the child. To allow the input to modify the innate schemas appropriately, one must specify how particular aspects of the input, such as the meanings of words, their linear sequence, and the predicate-argument relations perceived to be holding among them in a given situation, trigger the very first changes in the infant’s syntactic knowledge. One solution to this problem can be found in Macnamara’s (in press), Jackendoff’s (1977), and Grimshaw’s ( 1981) suggestions that syntax and semantics may correspond in
246
S. pinker
certain unmarked or canonical cases (e.g., physical objects are realized syntactically as nouns, human actors are represented syntactically as subjects, predicate-argument relations are realized syntactically as sister nodes in the phrase structure tree). If parental speech to children respects these canonical correspondences, it is fairly easy to show how a child could acquire her first phrase structure rules by using a combination of word order, semantic information, and innate constraints on phrase structure such as those found in X-bar theory (Pinker, 1981 b; Grimshaw, 1981; Berwick, 1980). At the same time, we have a learnability-theoretic explanation for why those unmarked correspondences should be so widespread in the world’s languages to begin with, and why syntax and semantics correlate highly (though not perfectly) in chiidren’s early utterances. Of course., it is imperative to show how the child acquires words and phrase structure rules that are never semantically transparent, but once some phrase structure rules have been acquired, this can be aczsmplished by a distributional analysis procedure examining the arrangement of unknown elements within a partially-constructed phrase marker (Grimshaw, 1981; Pinker, 198 lb). Finally, in a theory of grammar with a sharply restricted transformational component (e.g., Bresnan, 198 1; and possibly also Chomsky, 1979, and Gazdar, in press), it is also relatively straight forward to devise procedures to acquire many types of rules in the grammar once s3me phrase structure rules are in place,(Pinker, 198 1b; Berwick, 1980). An acquisition theory of this sort raises many empirical and theoretical questions, which I am now examining in collaboration with a number of colleagues (including ‘Jane Grimshaw, David Lebeaux, and Ani Zaanen). First, does parental speech to, ,rhildren exhibit the correspondences between syntax and semantics that the theory assumes? Second, by couching all of the partial knowledge states of the child in an autonomous syntactic vocabulary (albeit partially induced with the help of semantic input), can this theory provide stat&factory accounts of children’s linguistic abilities, such as the substitutability of lexical items of a given category within different phrase structures, and the domain of application of productive rules? We are initiating a program of developmental experiments designed to assess the form of syntactic rules such as passive, dative, and subject-verb agreement in preschoolers, to see to what extent the theory’s assumption of autonomous syntax in young children is viable. Finally, it is important to show how well the putative acquisition mechanisms work together when they are set to process complex linguist? input, and also how robust these mechanisms are in the face of inputs that violate the theory’s assumptions. It seems that computer simulation of the acquisition model (Walsh, 1981) will be highly useful in this evaluation. What, then, about the commonalities between visuospatial representation and syntax acquisition? Many people are disappointed when I conclude that
Spatial representation and lungu~e acquisition
247
in fact these abilities have little in common of any interest (i.e., apart from both components interfacing with semantic knowledge, or recruiting atomic information processes such as searches and comparisons). IFmy tentative proposals are correct, each ability seems to involve computational mechanisms that are highly specific to the tasks they must perform (such as the dual-address array, or the phrase structure rule schemas), and I find myself more in sympathy with Chomsky’s (1980) hypothesis of distinct “mental organs” than with proposals about a universal mental interlingua of elementary information processes. This conclusion should not come as a surprise wh,n one considers the full range of relievant data (including experimental, observational, intuitional, and computational considerations), and when one’s goal is an explicit computational theory of an ecologically significant cognitive ability. Perceiving and imagining a three-dimensional world which reveals itself through a two-dimensional retinal surface, and inducing the syntax of a natural language on the basis of word strings per,eived in situations, turn out to be exquisitely complex tasks but ones that our speci:s has evolved to compute with astonishing ease. I find it neither surprising nor disappointing to conclude that visual space and language acquisition use special-purpose co,gnitive structures and processes, and if I was forced to make a prediction about progress in the coming years, it would be that computationally explicit and empiricallymotivated theories of other cognitive faculties will come to that conclusion as well. References Anderson, J. R. (1977) Induction LP)I augmented transition networks. Cog. SC&I, 125-157. Attneave, F. (1972) Representation of physical space. In A. W, Melton and E. J. Martin (eds.), Coding processes in’human memory. Washington, DC, Winston. Attneave, F. (1974) How you know? Amer. Psychol., 29,493-499. Baker, C. L. and McCarthy, J. (1981) The logic01problem of longxage ocqukr’tion.Cambridge, Mass, MIT Press. Berwick, R. C. (1980) Learning structural descriptions of grammar rules from examples. MIT Intelligence Report #578, Cambridge, Mass. Bresnan, J. W. (1981) (ea.), The mental representation of grammoticu; reiations. Cambridge, Mass., MIT Press. Brown, R. (1973) A first Irmguage: The e&y stages. Cambridg.., Mass: Harvard University Press. Chomsky, N. (1962) Explanatory models in linguistics. In E. Nqel and P. Suppes (eds.), Logic, m&hodology ond philosophy of science. Chomsky, N. (1979) Markedness and core grammar. Unpublished manuscript, Mid,Cambridge Mass. Chomsky, N. (1980) Rules and representorions.New York, Columbia University Press. Gazdar, G. J. M. (In press) Phrasestructuregrammar.In P. Jacobson and G. K. Pulhlm (eds.), The nuture of syntacticrepresentation. Gibson, J. J. (1950) 7Yreperceptfon of the Visa*alword. Boston: Houghton Mifflin.
248
S. pinker
Grimshaw, J. (1981) Form, function, and the knguage acquisition device. In C. L. BJ I %qdJ. McCarthy (eds.), 7ZreZogi& problem ~fliwaguageacquisition Cambridge, Mass., MIT Fiess. Hinton,G. E. (1979) Imagery without arrays. Behuv. Br. Sci., 2,%X-556. lackendoff, R. (1977) X-barsyntax: A study of phase structure Cambridge, Mass, MLTPress. Kosslyn, S. M. (1980) Image end mid. Cambridge, Mass. Harvard University Press. Kosslyn, S. M. and Shwartz, S. P. (1977) A simulation of visual imagery. Cog Sci., I, 265295. Macnamara, J. (In press) NQWS for things. Montgomery, VT: Bradford Books. Marx, D., and Nishihara, H. K. (1978) Representation and recognition of the spatial organization of threedimensional shapes. Proc. R. Sot., 200,269-294. Phrker, S. (1979~) Formal models of Language Learning. Cog., 7, 217-283. Piuker, S. (1979b) Mental maps, mental images, and intuitions about space. Behav. Br. Sci., 2, 513. Pinker, S. (198Oa) Mental imagery abd the third dimension. J. exper. PsychoL, Z09,354-371. Pinker, S. (198Ob) Mental imagery and thevisual world. (Center for Cognitive Science Occasianal Paper ##4).MIT, Cambridge, Mass. Pinker, S. (1981a) The mental representation of 3-D space: A hypothesis. Unpublished manuscript, Harvard University. Pinker, S. (1981b) A theory of the acquisition of lexical interpretive grammars. In J. Bresnan ted.), llre men&drepresentationof grammutidreiktions. Cambridge, Mass, MIT Press. Pinker, S. and Fir&e, R. A. (1980) Emergent twodhnensional patterns in hages rotated in depth. J. exper. .Psychol.,Hum. Percep. Perf., 6. 244-264 Pinker, S. and Kosslyn, S. M. (1978) The representation and manipulation of threedimensiona! space in mental images. J. ment. Zmag..,2,69-84 Posner, M. I. (1980) Orienting of attention. Q. J. exper. Psychol., 32, 3-25. Shepard, R. N. (1978) The mentalimage.Amer. Psychol., 33,125-137. Shepard, R. N. and Metzbr, J. (1971) Mental rotation of three-dimensional shapes. Sci., Z71, 701-703. UBman, S. (1979) ;Ihe ik+terpretation of visualmotion. Cambridge, Mass, MIT Press. Walsh, R. (1981) A computer simulation of the acquisition of lexical grammars. Unpublished senior thesis, Harvard University, 1981. Wezler, K. and CuBcover, P. (1980) Formal principles of languageacquisition,Cambridge, Mass, MIT Press.
Cognitim, 10 (1981) 249-259 @ Elsev,er Sequoia S.A., Lausanne - Printed in The Netherlands
249
Some current theoretical issuesin speechperception DAVID
B. PISONI”
lndiam University
The field of speech perception is a very rich interdisciplinary area involving researchers from a number of disc&lines including psychology, lingui: +ics, speech and heariig science, electrical engineering and artificial intelligence. AS a consequence, the particular issues under study are often approached in different ways by different research groups. Despite these sulperficial differences, however, there are a small number of bade questions th,at can be identified as the core of problems that people are ~:~orkingon today. In my view, the fundamental problem is speech perception is to describe how the listener converts the continuously varying acoustic stimulus produced by a speaker into a sequence of discrete linguistic units and how he recovers the intended message. This problem can be broken down into a number of more specific subquestions. For example, what sta.ges of perceptual analysis intervene beZween presentation of a speech signal and eventual understandi:lp of the message? And, what types of operations occur at each of these stages? What specific types of perceptual mechanisms are involved in speech perception and how are they used in understanding spoken language? In understanding spoken language we CIi.ssume that various types of information are computed by the speech processing mechanisms. Some forms of information are transient, and last for only a short period of time; other forms of information are more durable, and interact with other sources of information from long-term memory (see Pisoni, 1978). The nature of these perceptual codes, and their interactions are among the major concerns in the field of speech perception today. In the remainder of this paper I wi!! briefly review what I see as the major theoretical issues in speech perception. Some of these have been discussed in the past and continue to occupy a central role in speech perception research today; others relate tar new problems to be pursued in the future. Linearity, lack of invariance anal the wgmentation problem For more than thirty years, it has been extremely difficult to identify acoustic segments and features which uniquely match the perceived phonemes inde*This research was supported by NIHresearch grant NIS12179 to Indiana University in Bloomington. I an’grateful to Jon Allen of MIT for many fruitful conversations regardingthe issues dircussed in this paper. Reprint requests should be sent to: David B. Pir~&, Speech Research Laboratory, Department of Psychology, Indiana University, Bloomington, Indiana 47405, U.S.A.
pendencly of the surrounding context. Often a single acoustic segment contains information about several rieighboring linguistic segments (i.e., parallel transmission), and, conversely, the same linguistic segment is often represented acoustically in quite different ways depending on the surrounding phonetic context, rate of speaking and talker (i.e., context conditioned variation). In addition, the acoustic characteristics of individual speech sounds and words exhibit even greater variability in fluent speech because of the influence of the surrounding context. The context-conditioned variability in the correspondence between acoustic signal and phoneme alzo presents enormous problems for segmentation of speech into phonemes or even words. Because of the failure to meet the linearity and invariance conditions, it has been difficult to segment speech into acoustically defined units that are independent of adjacent segments or are free from contextual effects of sentence contexts. It has been difficult to determine strictly by simple physical criteria where one word ends and another begins, especially in connected speech. Although segmentation is possible according to strictly acoustic criteria (see Fant, 1962)., the number of acoustic segments is typically greater than the number of phonemes in the utterance and, moreover, no simple invariant mapping has been found between these acoustic attributes and perceived phonemes or individual words. Continued attention will no doubt be directed at describing the decoding strategies used by listeners in interpreting contextdependent acoustic pattrsrns that correspond to phonemes in words. Other research will be concerned with how these cues are modified b_ythe acoustic characteristics of sentences in fluent speech.
Internal
representation of speech signals
Until very recently there was fairly gocd agreement among investigators working on human speech perception that at some stage of perceptual processing speech is represented internally as a sequence of discrete segments and features (see Studdert-Kennedy, 1974, 1976). There was less agreement, how ever, about the exact description of these features. Arguments were pro@! ;d for feature systems based on distinctions in the acoustic domain, the aruculatory domein, and systems which combine both types of distinctions. Recently there has been a trend to view these sorts of traditional featurn descriptions of speech sounds with some skepticism, particularly with r &ai; to the role these features play in on-going perception (Ganong, 1979; Klatt, 1977, 1979; Raker, 1977). On reexamination, much of the original evidence cited in support of feature-based processing in perceptual experiments seems arnb&uous and equally consistent with more parametric representations of
Some current theoretical issues in speech perception
251
speech. As a consequence, numerous investigators have begun to look more closely at how speech waveforms are processed in the peripheral auditory system, and what thzse more detailed representations may contribute to questions surrounding phonelicprocessing and particularly the problem of acousticphonetic invariance (Klatt, 1979; Searle et !z1., 1979; Zwicker et a!. , 1979). One of these approaches has been to repr:sent speech in the frequency domain as a sequence of magnitu0.e spectra sampled about every 10 ms. or so. If these spectral samples are adjusted to take account of certain psychophysical facts about hearing (such as critical bands, spread of masking and the growth of ioudness), a continuous representation can be obtained that is similar to a ‘neural spectrogram’. Efforts along these lines have directed attention to the problem of tne neural representation of speech, and So questions having to do with psychological filtering. Very little is currently known about the way speech is represented at these early stages of processing. There are, of course, num.erous different ways of representing the perceptual dimensions of speech signals and these can be examined with regard to questions dealing with how a listener identifies the acoustic cues to various segmental contrasts.‘Moreover, there is the additional and quite separate problem Iof defining the proper psychological dimensions that can be used to characterize appropriate display.. of these dimensions. The peripheral auditory system can be modelled as a frequency analyzer with continsusly varying neural signals as output. Thus, the perceptual dimensions and the ‘neural’ rep resentations of speech should be based on what is currently known about the function of the peripheral auditory system and how it processes various types of acoustic signals. Normalization problem In addition to the problems arising from the lack of acotlstic-phonetic invariance discussed earlier, there are also two distinct problems having to do with normalization of the speech signal. One is the talker-normalization problem. Talkers differ in the length and shape of their vocal tracts, in the articulatory gestures used for producing various types of phonetic segments, and in the types of coarticulatory strategies present in their speech. As a consequence, there are very substantial differences among talkers in the absolute values of the acoustic correiate; of many phonetic features. Differences in stress and speaking rate as well as in dialect and affect also contribute to differences in the acoustic manifestation of speech. Clearly, the invariant properties cannot be absolute physical values encoded in the stimulus but instead must be relational in nature. Unfortunately, there is relatively little known at present about
252
David B. Pisoni
this form of perceptual normalization or about the types of mechanisms involved (however, see Remez et al., 198 1). The se’cond problem concerns time and rate normalizaticin. It is now well known that the durations of individual segments are inAenced quite substantially by an individual talker’s speaking rate. However, the acoustic durations of segments are also affected by the locations of various syntactic boundaries in connected speech, by syllabic stress, and by the component features of adjacent segments in words ((seefor example Gaitenby, 1965 ; Klatt, 1975, 1976, 1979; Lehiste, 1970). In addition, there are substantial differences in the durations of segments of words when produced in sentence contexts compared to the same words spoken in isolation. The rate at which speakers talk also influences the duration and acoustic correlates of various phonetic features and segments. Numerous low-level phonetic and phonological effects such as vowel reduction, del.;tion and various types of assimilation phenomena have been well documented in the literature. These effects are influenced a great deal by speaking tempo, dialect and surrounding phonetic context. It has also been known for many years that duration can also be used to distinguish various segrn:ntal contrasts. Many phonetic and phonological segmental contrasts are also distinguished by redundant differences in duration as well as by their primary spectral correlates. Thus, the listener is faced with the problem of trying to ignore certain kinds of irrelevant durational information while trying to incorporate other kinds of distinctive durational information about segment,s, stress, prosody and syntactic s’tructure (see Miller, 198 1; Port, 1981 for reviews). Units in speech perceptian Another important and long-standing issue in speech perception is the choice of a minimal unit of perceptual analysis. Because of limitations of channel capacity, especially in the auditory system, raw sensory information must be recoded into some more: permanent form that can be used for subsequent analysis. Is there a basic or ‘natural’ coding unit for speech perception? Many investigators have argued for the primacy of the feature, phoneme, syllable or word as their candidate for the basic perceptual unit. Other investigators have even proposed larger units for perceptual analysis of speech, such as clauses of sentences @ever, Lackner and Kirk, 1969; Miller, 1962). The debate over the choice of a perceptual unit can be resolved if a strict distinction were made concerning the level of linguistic analysis under consideration. The size of the processing unit in speech perception varies from feature to segment to
Some current theoretical issues in speech perception
253
over the question of whether there is one basic or primary unit are inappropriate since there are, in fact, many units that are used by the speech processing mechanisms. Prosody, rhythm and speech timing Most of the research in speech perception over the last thirty years, as well as the major theoretical emphasis, has been concerned with segmental analysis of phonemes. One seriously neglected topic has been the prosodic or suprasegmental attributes of speech, which involve differences in pitch, intensity, duration, and the timing of segments and words in sentences. At present there remains a wide gap between the research on isolated segments and features and prosodic factors (see Cohen and Nooteboom, 1975). It is clear, however, that this source of linguistic information serves to link phonetic segments, features and words to grammatical processes at higher levels of analysis (see Darwin, 1975; Huggins, 1972; Nooteboom et al., 1978 for reviews). Moreover, speech prosody may also carry useful i.nformation about lexical, syntactic and semantic properties of the speaker’s message. It would be of interest to know, for example, the extent to which syntactic and semantic variables influence the durations of phonetic segments ancl words, and whether listeners iarr and do use this sort of information in understanding spoken language (see Hug,gins, 1972, 1978; Klatt and Cooper, 1975).
Lexical access and word recognition The problems of word recognition and the nature of lexical representations have been long-standing concerns of cognitive psychologists, although these problems have not been studied ex.tensively by investigators working in the mainstream of speech perception. This is true because the bulk of work on word recognition was concerned with investigating visual processes with less attention directed to questions of spoken word recognition Moreover, most of the interest in speech perception has been directed toward feature and phoneme perception which typically used isolated nonsense syllables. While such an approach is appropriate for studying ‘low level’ acoustical analysis of speech, it is not very helpful in dealing with questions surrounding how meaningful words are recognized in isolation or in connected speech. There are several interesting and important problems in speech perception that touch intimately upon the process of lexical access and bear more directly on the nature of the various types of representations in the mental lex-
254
David B. pisoni
icon. For example, it is of considerable interest to determine precisely what kinds of representations exist in the mental lexicon. Do words, morphemes, phonemes, or sequences of spectral templates represent lexical entrie:;? Is a word accessed on the basis of an acoustic, phonetic or phonological code? Are high frequency words recognized more-or-less automatically by very rapid search through a special precompiled network’? Are less frequent words analyzed by general rules for morphological analysis? One of the central problems .in word recognition and lexical access deals with the interaction of sensory input and higher-level contextual information. Some investigators, such as Forster ( 1976) and Massaro ( 1977), maintain that early sensory information is processed independently of higher-order context, and that the facilitation effects observed in word recognition are due to postperceptual processes involving decision criteria. Other investigators such as Morton (1969, 1979), Marslen-Wilson and Welsh (1978), Marslen-Wilson and and Tyler ( 1980), Cole and Jakimik ( 1978) and Foss and Blanck ( 1980) argue that context can, in fact, influence the extent of’early sensory analysis of the input signal. Klatt (1979, 1981) has recently proposed a model of lexical access t+iat explicitly avoids any need to compute a distinct level of representation corresponding to a sequence of discrete phonemes. Instead, he has precompiled an abstract phonetic lexicon of all possible words into a network of sequences of spectral templates. These templates are context-sensitive much Bikethe earlier ‘Wickelphones’ (Wickelgren, 1976) since they are supposed to characterize the acoustic correlates of phones in different phonetic environments by encoding the spectral characteristics and transitions from the middle of on& phone to the middle of the next. Klatt (1979) argues that this form of diphone concatenation is sufficient to capture much of the context-dependent variability observed for phonetic segments in spoken words. Much remains to be done to access these claims as valid psychological descriptions of the representation of words in the mental lexicon. Phonetic and phonological recoding of words in sentences One of the major difficulties encountered in speech perception is that each utterance of a language can be realized phonetically in many different ways. Obviously, it is unrealistic to store every possible utterance of the language in long-term memory Gnce the number of different sentences and phonetic realizations is potentially infinite. While it might be possible to adopt this strategy in the case of machine recognition of speech in very limited corltext, such a strategy seems inappropriate in the case of human speech perception.
Some current theoreticalissuesir speech perception 255
In addition to general phonological processes which characterize certain uniform dialect differences in pronunciation among talkers, there are also sets of low-level phonetic implementation rules which characterize more specific acoustic-phonetic variations among individual talkers. Because the number of different phonological phenomena in language is quite large, and because of the idiosyncratic variability of individual talkers, sets of decoding rules have been formulated from careful study of the acoustic and phonetic properties of speech in various contexts. Rules such as these must also be assumed to be part of the perceptual strategies used by human listeners in understanding spoken language, Despite the long-standing interest in phonological processes by linguists and the importance they play in the acoustic-phonetic realization of spoken language, relatively little perceptual research has been directed toward these problems. With the use of synthesis-by-rule systems, sets of phonological and phonetic implementation rules can be formulated and the effects of variations and modifications in these rules can be studied with isolated words and words in sentence contexts (see Huggins, 1978). Focused search and ‘islands of reliability’
There can be little doubt after some thirty years of research on speech that the acoustic signal contains a great deal of redundant information. A basic engineering goal has been to try to locate the most important information and code it in the most efficient way for trammission. In the same way, investigators concerned Nith human speech perception have tried to identify the ‘minimal cues’ for phonemes in the hope that once these could be identified the basic problem of recognition of speech could be solved. Unfortunately, there is a great deal more to speech perception and spoken language understanding than simply discovering the minimal cues for phonemes. The speech signal appears to be rich with salient and reliable inform;ition that listeners use in unders,tanding the message. As a consequence, the basic problem becomea one of finding in the stimulus ingut these ‘islands of reliability’ that can be used to access various different sources of knowledge. The term focused search has been used to characterize the strategies that listeners or intelligent machines use for inspecting the signal for information that can be ,useful at any given point in the perceptual process; focused search also specifically avoids information that does not provide useful support. Examples of such reliable information include: thd presence of stressed syllables, the beginnings and ends of words, and the locations of various spectral changes indicating shifts in the source function.
256
LbvidB. pisorli
Focused search emphasizes an important problem, namely, to identify those acoustic correlates of the signal that the listener relies on. The scope of a listener’s focus& seuch strategies varies substantially with the requirements of experimental tasks; what may be salient and reliable acoustic-phonetic information in one lis?ening cantext may not be used at all in another. Research on this problem has shifted recently from experiments using isolated nonsense syllables which are manipulated in very precise ways to investigations directed at ho-w listeners use these cues to perceive words in isolation and in sentence contexts whelre several diverse sources of knowledge can be used. The principle of delayed binding Human speeclh perception and spoken language understanding take place very rapidly in real time although relatively Yitcleis currently known about the processes and1 mechanisms that support z;uch on-lirlc activities. A good deal of the speech perception process occurs automatically and is therefore unavailable for oirect conscious introspection. Do all decisions at all levels of the speech perception process take place immediately in real-time or are there selected processing delays at particular analytic levels pending additional information? What is the size of the scanning window over which low-level phonetic dezisions are made? What depth of Iprocessing is required before a final and binding decision Carl be made about the se¥ta1 composition of the input signal? These are questions that are being pursued at this time by a number of researchers. The ‘principle of delayed binding’ evollved from the ARPA speech understanding project (see Klatt, 1977 for a review). According to this principle, decisiolns at 10w levels of processing are not forced if the information is unreliable or in:sufficient to make a final decision (see also Miller, 1962). Of course, such a priiciple might be appropriate in computational situations where the front-end or basic acoustic-phonetic recognition device fails to perform as well as humans do. But we know from much of the earlier research on word intelligibility that human listenerls can and do make binding low-level segmental and lexical decisions with extremely high accuracy even under very poor listening conditions. After alI, if the quality of the acoustic-phonetic information is very good, as in high-quality natural speech, phonetic, lexical and even syntactic decisions can occur on-line quite rapidly. However, in situations where the speech signal is physically degraded or impoverished, the speed of perceptuaI processing may be substantially slower, and certain low-level decisions may well have to be delayed pending higher-order constraints (Miller, Heise and Lichten, 195 1; Miller and Isard, 1963). In future research, it will
Some current theoretical issues in speech perception
257
be important to find out more about the perceptual and interpretative processes in human speech perception that aire responsible for the very rapid processing and the seemingly immediate on-line interpretation of spoken language as it is heard. Conclusion
The bulk of research on speech perception over the last thirty years has been concerned principally, if not almost exclusively, with feature and phoneme perception in isolated contexts using nonsense syllable materials, This research strategy has undoubtedly been pursued because very substantial problems arise w6en one deals wilh issues such as spoken language understanding and the relationship between ear& sensory input or the problem of word recognition and its interface with higher levels of linguistic analysis. Researchers in any field of scientific investigation typically work 011tractable problems that can be studied with existing methodology and paradigms. Relative to the bulk of speech perception restiarch on isolated phoneme perception, very little is actually known today about how the early sensory-based acoustic-phonetic information is used by the speech processing system in tasks involving word recognition and sentence perception or how changes in the segmental and/or suprasegmental structure of the speech. signal influence intelligibility and comprehen,sion of spoken language. These are problems of current interest that will no doubt 5e pursued over the next few years. I believe that continued experimental and theoretical work in speech perception will1be directed at new models and theories that capture significant aspects of the process of spoken language understanding. What is important at the present time, is to diret;i: research efforts toward somewhat broader issues involving the use of meaningful stimuli in more naturalistic experimental tasks that require the listener’s active deployment of phonological, lexical, syntactic and semantic knowledge to assign an interpretation to the sensory input. Past theoretical work in speech perception has not been very well developed nor has the link between theory and empirical data been very sophisticated. Moreover, work :n the field of speech perception as in other areas has tended to be defined by :zdecific experimental paradigms or particular phenomena (i.e., dichotic listening, categorical perception or selective adaptation). The major theoretical issues in speech perception often seem to be ignored. They receive little serious attention by investigators who are involved in working on the details of experhnczltal problems that unfortunately’ bear only marginally on the primary perceptual and cognitive processes that are used in spoken language understanding. Although a very formidable task, research
in the future will be focused more directly on the general problem of spoken Iangaage understanding. In my view, it is here that the greatest insights into language processing will be found in the next ten years. References Bevcr. T. G., Lackner, J., a;:d Kirk, R. (1969) The underlying structure sentence is the primary unit of immediate speech process~mg.Percep. PWychophys., 5, 225 -234. Cohen, A., and Nooteboom, S. (eds.). (1975) Smcture and Process in Speech Perceprion. Heidlelberg, Springer-Verlag. -Cole, R. A., and Jakimik, J. (1978) Understanding speech: How words are heard. In G. Underwood (ed.), Strategies of Informationfiocesssig. New York, Academic Press, pp. 68-l 16. Darwin, C. J. (1975) On the dynamic use of prosody in speech perception. In A. Cohen and S. G. Nooteboom (eds.), Strucrthrre and Process in Speech Perception Berlin, Springer-Verlang, PP. 178-194. I$nt, C. G. M. (1962) Description analysis of the acoustic aspects of speech. Logos, 5, 3-17. Forster, K. I. (1976) Act l ssing the mental lexicon. In R. J. Wales and E. Walker (eds.), New Apprwches to LongrtageMechanisms Amsterdam, North-Holland, pp. 257 -? 87. Foss, D. J.,and Blank, MM. A. (1980) Identifying the sreer’ codes. Cog. PsychoZ.,22, l-31. Gaitenby, J. H. (1965) The elastic word. In HaskinsLaboratories StatusReport on Speech Research, SR-2, 3.1-3.12. Gmong, W. F. (1979) The internal structure of consonants in speech perception: Acoustic cues, not distinctive feaiures. Unpublished manuscript. Hums, A. W. P. (1972) On the perception of temporal phenomena in speech. J. acoust. Sot. Am., 51, 1279-i 290. Huggins, A. W. F. (1978) Speech timing and intelligibility. In J. Requin (ed.), Attention and Perfcrmance VZ!.Hillsdale, NJ, Erlbaum. Klatt, D.. H. (I 975) Vowel lengthening is syntactically determined in a connected discourse. J. Phon., 3.129-140. Klatt, D. M. (1976) Linguistic uses of segmentaldurationin English: Acoustic and perceptual evidence. J. acousf. Sot, Am, 59, 1208-1221. Klatt, D. H. (1977) Review of the ARPA speech understanding project. J. acoousr.Sot. Am., 62,13451366. K&t, D. H. (1979) Speech percetion: A model of acoustic-phonetic analysis and lexical access. X man, 7. 279.-312: Wit, D. H. (1979) Synthesis by rule of segmental durations in English sentences. In B. Lindblom and S. Ohman (eds.), FIntiers of Speech Communication Research. New York, Academic Press. Klatt, D. H. ( In press) Lertcal representations and processing strategies during speech production and perception. Psychol. Ret Watt, D. H., and Cooper, W. E. (1975) Perception of segment duration in sentence contexts. In A. Cohen and S. G. Nooteboom (eds.), Structure and Process in Speech Perception, New York, Springer-Verlag. IlehisTe,1. (1970j SuprasegmerzfalsCambridge, MA., MIT Press. Marslen-Wilson, W. D., and Tyler, L. K. (1980) The temporal structure of spoken language understanding. cog., 8, l-71. Ma&n-Wilson, W. D., and Welsh, A. (1978) Processing interactions and lexical access during word recognition in continuous speech. Cog, Fsychoi., 10, 29-63.
Some current theoretical issues in speech perception
259
Massaro, D. W. (1977) Reading and listening. Technical Report No. 423, Wisconsin Research and Development Center for Cognitive Learning, University of Wisconsin-Madison. Miller, G. A. (1962) Decision units in the perception of speech. IRE Transactions on Information Theory, IT-S, 81-83.
Miller, G. A., Heise, G. A., and Lichten, W. (1951) The intelligibility of speech as a function of the cantext of the test materials. J. exper. Psychol., 41, 329-335. Miller, G. A., and Hard, S. (1963) Some perceptual consequences of linguistic rules. J. verb. Learn. verb. Behav., 2, 217-228.
Miller, J. L. (1981) The effect of speaking rate on segmental distinctions: Acoustic variation and perceptual compensation. In P. D. Eimas and J. L. Miller (eds.), Perspectives on the Study of Speech, _Hillsdale,NJ, Erlbaum Associates. Morton, J. (1969) Llteraction of information in word recognition. Psychol. Rev., 76, 165-178. Morton, J. (1979) Word recognition. In J. Morton and J. D. Marshall (eds.),PsychoZinguistics 2: Structuresand Processes. Cambridge, MIT Press, pp. 107-156. Nooteboom, S. G., Brokx, J. P. L., and deRooij, J. 1. (1978) Contributions of prosody to speech perception. In W. J. M. Levelt and G. B. Floresd’Arcais (eds.), Studies in the Perception of Language. New York, John Wiley, pp. 75-107. Parker, F. (1977) Distinctive features and acoustic cues.J. acoust. Sot. Am., 62, 1051-1054. Pisoni, D. B. (1978) Speech perception. In W. K. Estes(ed.)., Handbook of Learning and Cognitive Processes (vol. 6) Hillsdale, NJ, Erlbaum Associates, pp. 167-233. Port, R. F. (1981) Combinations of timing factors in speech plvd,ction. J. acoust. Sot. Am., 69, 262-274. Remez, R. E., Rubin, P. E., Pisoni, D. B.,andCarrell, T. D. (19Pl) Speech perception without traditional speech cues. Science, 212, 947-950. Searle, C. L., Jacobson, J. Z.,and Raymcnt, S. G. (1979) Stop consonant discriminatijn based on human audition. J. Acoust. Sot. Am., 65.799 - 809. Studdert-Kennedy, 51. (-974) The perception of speech. In T. A. Sebeok (ed.), Current Trends in L.‘yguistics (vol. XI.). The Hague, Mouton. Studdert-Kennedy, M. (1976) Speech perception. Hn N. J. Lass (ed.), Contemporary Issues in ExperimentalPhonetics. New York, Academic Press, pp. 243-293. Wickelgren, W. A. (1976) Phonetic coding and serial order. In E. C. Carterette and M. P. Riedman (eds.), Handbook of Perception (vol. VMI). New York, Academic Press, pp. 227-264. Zwicker, E., Terhardt, E., and Paulus, E. (1979) Automatic speech recognition using psychoacoustic mode1s.J. acoust. Sot. Am., 6.5,487-498.
Cognition, IO (1981) X1-266 @Msevier Sequoia S.A., Lausanne - Printed in The Netherlands
261
Cognjtion and neural systems MICHAEL I. POSNER* Univtkity
of Oregon
The past twenty years of research in the field of attention has witnessed a steady trend toward the study of internal covert mechanisms that lie behind the performance of skilled tasks. My own work has reflected this trend by starting with studies or okilled human perfomance, shifting to explorations of cognitive structure and later to model tasks that allowed more contact with neural systems. Human performance
The term human performance was introduced two decades ago to describe an approach that sought direct studies of everyday life skills. Research in the area emphasizes quantitative treatment of input output relationships. This levei of analysis is often most appropriate to application in a variety of military and industrial settings (Fitts and Posner, 196’7j. Cognition
The field of cognitive psychology (Neisser, 1967; Posner, 1973) reflects a greater emphasis upon internal mental structures and transformations that lie between input and output. A major area of cognitive research has been the study of speech perception and reading. The roles of automatic and conscious processes in producing sensory specific, phonetic and semantic codes have been examined in detail (Posner, 1978). This work has given us new perspective on the breakdown of cognitive function in brain injury (Coltheart, Patterson and Marshall, 1980) and upon its normal development over the lifespan (Posner and Rothbart, 1980. *This research was supported by NSF Grant BNS 792537 to the University of Oregon. Portions of this paper were presented to the 39th meeting of the Neuroscience Research Associates. I wish to thank many colleagues who have been involved in the research program described in this paper and Mary IL Rothbart for her help in clarifying this work Reprint requests should be addressed to M. I. Posner, Psychology Department, University of Oregon, Eugene, Oregon 97403, U.S.A.
262 M I. Posner
C0gnitiwneuroscience
As our understanding of the role of internal mechanisms in producing complex behavior has increased, cognitive psychologists have quite naturally cozze into greater contact with neuroscience disciplines. While the goals of neuroscience are focussed on identifying general principles of nervous system structure and function, I believe the methods and ideas of cognitive psychology can play an important role in helping to formulate these principles and in applying them to the problems of cognition. Such problems include both the study of uniquely human processes such as reading and also more general cognitive abilities that we share with other organisms. E$atid attention
During the last several years my colleagues and I have sought to foster the connections between cognition and neuroscience through the study of spatial attention (Posner, 1978; Posner, Pea and Volpe, in press). We have tried to devellop paradigms for the study of attention in normal human beings that would make contact with developing neuroscience studies of attention using single cell recording from alert monkeys. In an effort to get beyond demonstrations that models of cognition can be loosely related to problems of brain injury, we have attempted a more detailed analysis of hypotheses arising in both neuroscience and psychology. For example, there has been active interest in the relationship between attention and movement in both neurophysiology (Goldberg and Wurtz, 1972; Mountcastle, 1978) and in cognitive psychology (Posner, 1980). For visual events, the major interest has been in the relationship between orienting (overtly by eye movements, or covertly via shifts of attention) and the efficiency of detecting (making arbitrary responses, or being aware of) stimuli. In our behavioral work, we have been able to explore three general points: 1. Meastirement of covert orienting of attention by changes in the efficiency of detecting stimulus events at different spatial positions. 2. The relationship between movements of covert attention and movements of the eyes. 3. The pathways controlling both covert and overt orianting. Measurement ofcovert attention We have a variety of methods (e.g., reaction time, probability of reporting near threshold stimuli) to measure the efficiency of detecting information at various positions in the visual field. Subjects maintain fixation, but if cued to
Cognition and neural system
263
shift attention to places other than the line of sight they are able to do so (Posner, Nissen and Ogden, 1978). Such shifts are accompanied by improved efficiency of performance in terms of the latency of responding to signals that occur at the expected position in comparison to those that occur at the unexpected position (Posner, Nissen and Ogden, 1978). We have shown that in the case of luminance detection, the fovea plays no special rule in the control of attention, although attention cannot be used to compensate for the fovea when acuity is important (Posner, 1980). Shifts of covert attention can be time-locked so that the changes in efficiency can be traced dynamically as attention is moved across the visual field (Shuhnan, hzrungton and McLean, 1979). Attention aud eye movement The time-locking of attention shifts to external signals allows testing a number of theoretical positions about the relationship between the position of the eyes and orienting of covert attention. We have shown that the occurrence of a peripheral event leads to a shift of covert attention to the area of the target about 150 msec prior to an eye movement (Posner, 1980; Remington, 1980). This occurs even when the subject has a strong incentive to maintain attention at fixation and its time course resembles that of the selective enhancement of superior colliculus units (Goldberg and Wurtz, 1972). If a central cue is use.d to instruct subjects to make an eye movement, no evidence for a shift in visual attention prior to the eye movement has been found (Remington, 1980). These findings (Posner, 1980) suggest that there are strong functional relationships between the shifts of visual attention toward the occurrence of peripheral stimuli but that there is no identity in the underlying physiological system. Nor can attention be viewed as closely coupled to the programming of the oculomotor system as propossd by efference thecries (Wurtz and Mohlea, 1976). The close functional relationship between attention movements and eye movements is similar to the relationship between eye movements and hand movements (Posner and Cohen, 1980). We have also explored the significance of the functional relation between attention and eye movements found in adults for understanding of the development of attentional mechanisms in newborns (Posner and Rothbart, 198 1). Pathways of control It has long been believed that the superior colliculus plays a special role in programming overt movements of the eyes. Mammals tend to have stronger
264
M. I. Power
pathways from the retina to the contralateral superior collicull!s than ipsilateral connections. We have tested the functional significance of this anatomical relationship by allowing our subjects to view stimulus displays monoc~larlyy (Posner and Cohen, 1980). Subjects are instructed to move their eyes in a natural fashion. When simultaneous events are presented on both sides of fixation, there is a strong tendency to move toward the temporal visual field in accordance with the anatomical connections cited above (Posner and Cohen, 1980). This asymmetry does not occur with eye movements to auditory commands nor does it occur strongly in conscious judgments of temporal order that do not involve movements of the ey-es. There appears to be a similar temporal bias in infants occurring even when only a single stimulus is preset ted (Lewis, Maurer and Milewski, 1979). Shulman (1979) sought to determine if a similar bias toward the temporal visual field existed in coverts shifts of attention. He used several methods to explore this situation. He first determined the advantage in reaction time when attention was brought to a position in the visual field by the occurrence of a single peripheral target. He had subjects view monocularly trials in which physical targets occurred simultaneously to the left and right of fixation. A bias toward the temporal visual field of the magnitude found with eye movements should have produced a temporal tield advantage of about 70 msec, but no such bias was found. A number of approaches using patients are now being explored to determine the role of the second visual system in the control of covert shifts of visual attention. Recent work by Holtzman, Sidtis, Volpe, Wilson and Gazzaniga (in press) using split brain subjects has shown that orienting cues given to one half brain can guide movements of the other half brain. This means that no*rcortical pathways play a role in the overall process of directing covert attention. Similarly, our results (Rafel, Posner and Walker, 198 1) with patients suffering from collicular lesions (progressive supranuclear palsy) suggest a role for these -nidbrain structures in the latency of covert orienting. Conclusions The work on reading and the study of spatial attention have employed a common set of approaches. Prominent among these is the detailed study of normal human subjects using various forms of mental chronometry, the use of selected populations to achieve insights into the formation and breakdown of cognitive function and the effort to relate results to perspectives developing from an increased understanding of nervous system function obtained from neuroscience stu&s. Yhese approaches maintain the empirical and analytic
Cognitionad neuralsys~.ws
255
style that has characterized the psychological study of adult cognition, but also help to ensure that studies remain in close contact with skills of daily life and with important problems of cognitive deficit. I believe the close contact with applied problems will continue to be important in my own future work. Although I can’t honestly indicate a dramatic breakthrough in following this approach, I do hope that the accumulating research findings an? the availability of new techniques will help to further illuminate the relationship between cognition and brain activity.
References Bushnell, M. C., Robinson, D. L. and Goldberg, M. I. (1978) Dissociation of movement and attention: neuronal correlates in posterior parietal cortex. A'eurosciences Abstracts, 4, 621. Fitts, P. M. and Posner, M. I. (1967) Human Performance. Belmont, Calif., Brooks Cole. Coltheart, M., Patterson, K. and ?larshall, R. C. (1980) Developmental Dyslexia. London, Routledge % Kegan Paul. Goldberg, M. E. and Wurtz, R. H. (1972) Activity of superior colliculus in behaving monkeys. II. ;,ftBct of attention on neuronal responses. J. Neurophysiol., 3.5 560-574. Holtzman, J. D., Sidtis, J. J., Voll:e, B. T., Wilson, D. H. and Gazzaniga, M. S. (In press) Dissociation of spatial information for stimulus localization and the control of attention. Brain. Lewis, T. L., Maurer, D. and Mile&ski, A. E. (1979) The development of nasal de&e&ion in young infants. ARVO Abstracts, ;rlay, p. 27 1. Mountcastle, V. B. (1978) E;ain mechanisms for directed attention. J, Royal Sot. Med., 71 14-27. Neisser, U. (1967) Co,@tiue Psycholggy. New York, Appleton-Century Crofts. Posner, M. I. (1973) Cognition: an introduction. Glenview, Ill., Scott, Foresman. Poner, M. I. (1978) Chronometric Exploration of Mind. Hillsdale,NJ, Lawrence Erlbaum Associates. Posner, M. I. (1980) Orienting of attention. The VIIth Sir Frederic Bartlett Lecture. Q. J. exper. Psychol., 32, 3-25. Posner, M. I. and ‘Cohen, Y. (1980) Attention and the control of movements. In G. E. Stelmach and J. Requin (eds.), ?Wonids in Motor Behavior. Amsterdam, North Holland Publishing Co., 243258. Posner, M. I., Nissen, M. J. and Ogden, W. C. (1978) Attended and unattended processing modes: The role of set for spatial location. In Pick, H. E. and Saltzman, I. J. @Is.), Modes of Perceivingand Processinginformation. Hillsdale, NJ, Lawrence Erlbaum Associates. Posner, M. I., Pea, R. and Volpe, B. (In press) Cognitive-Neuroscience: Developments toward a science of synthesis. To appear in Mehler, J. (ed.), Roceedings of the CNRS Conference, Royaumont, June X4-18,1980. Posner, M. I. and Rothbart, M. (1981) The development of attentional mechanism. In J. Plowers (ed.), The Nebraska Symposium 1980. Lincoln, University of Nebraska Press. Rafel, R., Posner, M. I. and Walker, J. (1981) Progressive supranuclear palsy: Clinical and experimental observations. Paper presented to International Neuropsychology Society, Atlanta, February. Remington, R. W. (1980) Attention and saccadic eye movements. J. exper. Psychoi. Hum. Percep. Perf., 6, 726-744.
266
M. I. Posner
Shuhnan, G. L, (1979)Spatial dete rminants of attention allocation. Unpublished doctoral dissertation,
univefsity of Oregon. Shu!man, G. L, Remingtcq R. W. and McLean, 1. (1979) Moving attention through visual space. J. expw. l?sychoLHum. Pemp. Perfi, $522426. Wurtz, R. H. and Mohler, C. W. (1976) Organization of monkey superior coUiculus: Enhanced visual response of supe&ial layer cells. J. NeumphysioL 39, 745-765.
Cognition, 10 (1981) 267-274 @ 3evier Sequoia S.A., Lausanne - Printed in The Netherlands
267
Psychological explanations and knowledge-dependent processes ZENON W. PYLYSHYN* University of Western Ontario
The experimental and theoretical research that I have been engaged in during the past decade has been guided by certain general views concerning the nature of explanation in cognitive psychology. It turns out, perhaps surprisingly, that some pretty fundamental FZlilosophical considerations make an immediate difference to one’s d-.y to day research. Psychology is like that: The distance between one’s philosophical views (whether they are implicit or explicit) and everyday psychological investigations is remarkably short. One has but to look at the introductory and concluding sections of reports of stmightforward laboratory observations to see that what experimenters think is’3eing studied and what morals they arti prepared to draw from the resuits is the,*oughly conditioned by the stand that they take on the most basic philosop+ ical questions about the nature of mind and of psychological explanation. The basic position that has guided my own research is this. It seems that there are two distinct kinds of processes responsible for our behavior, each of which requires a different form of explanation becacse its regularities are governed by quite different principles. The first kind of explanation (Type I) is what might be called a cognitivist or knowledge-based or rational account, which explains certain aspects of behavior as a rational consequence of th.e organism’s possessing specific beliefs and goals. The operative principle in this sort of explanation is generally, though not exclusively, that of rationality (though a rationality whose excercise is conditioned by various limits on processing capacity). In the simplest case we might say that an organism performs action A because it has goal G and believes that carrying out A will help achieve G. Such explanations are entirely routine in large segments of psychology (e.g., social psychology) where decision-making is considered a central process. In other areas, however, (e.g., learning) it is looked upon with suspicion and generally treated as merely an informal way of speaking, and as replaceable, at least in principle, by a more objective or nr?ralistic account. There is ve.ry good reason to believe, however, that this way of talking is essential: that without such terms as goals, tacit knowledge, utilitie:;, inferences, *Reprint requests should be sent to: Zenon Pylyshyn, Department of Psychology, The University of Wetexn Ontario, London, Ontario, Canada, N6A X2.
268
Zemn W.qYlyshyn
and so on, we would not be able to capture important generalizations about hum&n behavior. The second kind of process requires what might be called a naturalistic or mechanical-cau~ explanation (I will call this Type II to avoid evoking unfortunate prejudices). In this case the behavior of the organism is explained in terms of its possessing certain intrinsic properties which are lawfully connected in ways that cause it to behave as it dots. Although an ideally complete explanation of such processes might derive the properties and regularities in question from basic biological principles, in practice one is generally satisfied with a characterization of the regularities at a more global (or ‘functional’) level. One might usefully think of behaviors requiring Type II explanations as being in a sense ‘twired in’. Unlike the hard-wiring on a typical computer, however, the ‘wiring’ in the brain is relatively malleable. Despite the existence of such general neurological malleability, however, we need to distinguish the sorts of changes to an organism that arise as a rationally comprehensible consequence of its relations with other cognitive states or with a meaningful interpreted environment-say changes that take place when a person is toId something or sees or hearssomething which he or she interprets in some way, as opposed to those changes that take place only through non-rational means (e.g., maturation, psychosurgery, biochemistry, or perhaps rate repetition or even the operation of such principles as contiguity or reinforcement). The reason is that these two forms of change appear to follow quite different principles; specifically the first requires that we attend to the meaning, or the semantics, of the causal event, while the latter does not. The distinction b%ween these two types of processes arises in explaining certain regularities in behavior when the latter are characterized in terms of meaningful actions -actions such as, for example, answering questions, making assertions, reaching for the salt, buying a car, or any other activity whose charanterization involves essential reference to one’s intentions. One of the main reasons why this sort of behavior has to be explained in terms of goals, beliefs and inferences is that it can frequently be altered in a coherent and rationahy explicable marmer by merely providing the subject with certain information-independently of how and in what physical form tbat information is in fact transmitted. This observation is the basis of a useful methodological principle (called the ‘cogniti.ve penetrability’ condition) for decidmg whether a certain process is Type I IorType II. Suppose that the function (in sense of a mathematical input-output relation) carried out by some stage of a psychologicaI process can be shown to be freely manipulable by certain external effects. Suppose, further, that the nature of thevariation is such that it appears both systematic and coherently explainable only when the influencing effects are viewed as sigrmh to which the organism has given some
Psycholugicar!explanations and knowledge-dependent processes 269
particular interpretation. In such cases the function itself cannot be satisfactorily explained as a Type II process. It must, at least in part, be given a Type I account since clearly some interpretation-dependent rational process, like making an inference or a decision must be involved. For example, if the way people react to some stimulus can be radically changed by telling or shtiwing them something (thereby perhaps changing their goals’or their expectations or beliefs about the stimwlus, or changing the way that they interpret the task) then clearly there must be some stage of interpretation, inference or decision-making taking place in the overall process that determines the S-R function in questions. Thus if Brewer ( 1974) is right in claiming that virtually all cases of reported human conditioning are ‘cognitively penetrable’ in this sense -i.e., that the effects can be induced, systematically changed, or even eliminated entirely by merely informing subjects of the contingencies of the experiment-then at the very least the explanttion of these results must appeal to more than just the built-in mechaarism of conditioning. Some stage in the process must involve decisions based on what subjects believe, and hence requires a Type I explanation. I have also used this criterion to argue, for example, that certain proposals for ‘analogue’ processes in cognition are not well founded, or that the claim that phenomena such as those associated with mental imagery arise from certain built-in properties of a medium or a ‘surface display’ are incorrect. I will briefly mention one or two of these cases below. One of the important discoveries of cognitive science is that it is possible to reconcile Type I processes with a materialist view of causation, and thereby avoid the retreat to dualism. This is done by hypothesizing the existence of (1) symbol structures (i.e., representations) that encode beliefs and goals in some physical form, and (2) elementary operations, which are themselves presumeably Type II processes in the brain, but which can be arranged to carry out the required Type I processes-just as the operations built into a cornputer can be used to carry out rational processes like reasoning and decision making. Indeed, the distinction between the two principal kinds of processes we have been discussing is exactly parallel to the distinction drawn in cornputer science between the program and what is called the funct!onal architecture of the virtual machine, represented by the hardware and the resources provided by the programming language. My contention has beer; that drawing the distinction between Type I and Type II processes is of fundamental importance to the task of understanding cognitive phenomena. In the first place failure to distinguish between those properties of the overall process that arise from rules-and-representations and those that arise from biological properties of the underlying medium, allows one to construct ad hoc computational models that will mimic any observed
270
Zenon W.Pfvlyshyn
input-output behavior with very little constraint-leaving too many degrees of freedom to permit the claim that the resulting mdJde1is explanatory (see the discussion in Py13Vshyn, 1979b, 1980). In the second place, failure to make the distinction frequently leads people to treat the discovery of some behavioral regularity as a discovery about some property of mental mechanisms (Type II) whereas it should, in fact, be treated as a discovery about subjects’ tacit knowledge of certain properties of the real world, or as a discovery about what the subjects take their task to be (this issue, as it applies to the case of mental imagery studies, is discussed at length in Pylyshyn, 198 1). There are numerous examples reported in the literature which show that a Type I (representational) process is required to explain phenomena that people had interpreted as indicating certain basic properties of mind. I have already cited the example of human conditioning experiments, But one of the earliest victims of this demonstration of cognitive peiletrability was in psychophysics, where Swets, Tanner and Birdsall (196 1) showed that the psychophysical threshold Function was cognitively penetrable-that it could be manipulated by altering subjects’ perceived utilities. Similarly, the psychophysical function responsable for the multidimensional structure of a similarity space, as derived from similarity judgments, was shown to be cognitively penetrable by Shepard (1964), who found that which (if any) metrical structure one obtains depends on which aspects of a stimulus the subject chooses to attend to-and the latter, in turn, depends upon. the subject’s goals, expectations, and so on. The cognitive penetrability of a large part of perception (almost everything past figure-ground isolation and stereopsis) is also one of the principal arguments against Gibson’s theory of direct perception (see Fodor and Pylyshyn, 1981). Similarly, my early arguments against the vie*v that images are stored in memory (Pylyshyn, 1973, 1978) relied in part on the observation that the way in which such memories fail and the way in which they are accessed all indicate that they are conceptual, rather than geometric, in nature: that the processes that operate on such memories are of Type I (which is one reason why I still prefer to call them ‘propositional’despite the widespread misunderstanding engendered by the use of this term). More recently we have obtained experimental evidence showing that a number of image manipulation processes, assumed by many to be evidence for Type II (frequently called ‘analogue’l ) processes, are in fact cognitively ‘The term l analogue* means many things to many people.I have used it here and elsewhere to refer to Type II processes. In Pylyshyn (1981) I argue that this is the sense of the term that is in fact relevant in arguments about the nature of mental representation. I have no views concerning whether or not Type I processes are inherently discrete, though it should be recognized that the only form of symbol systems that we: understand well enough to use as the basis of a theory of mental representation are discrete ones-such as computer data structures or various logical calculi. ;
Psychological explanations and knowledge-dependent processes
271
penetrable. For example, the rate at which subjects appear to ‘mentally rotate’ figures in figure comparison tasks such as those studied by Cooper and Shepard (1973) (in which ‘rotation rate’ is defined in terms of the slope of the best fitting straight line relating reaction time and angle of misalignment of two figures being compared) was shown to depend upon the conceptual complexity of both the figures in question and of the post-rotation comparison task. This led us to conclude that at least some stage in the process must be of Type I (or non-analogue), and hence that a holistic rotation account is incorrect (Pylyshyn, 1979a). In addition, a large number of mental image manipulation tasks such as those studied by Kosslyn (1980) (e.g. ones involving recording the relation between reaction time and the distance traversed in shifting one’s attention over a mental image, or the relation between time to extract informat. 3r-rfrom an image and the reported ‘size’ of the image) may also be ceg-.ttively penetrable. Thus in a number of separate studies we have shown that the linear relation between image distance and reaction time can be eliminated if the subject understands the task to.be something other than the deliberate simulation of certain properties of viewing a real physical event. This has led. us to suppose that it may be the case that in most such s+TAGes, what we are learning is that subjects have tacit knowledge of the real situation they are asked to imagine and that they have the psychophysical ability to generate the appropriate simulated results (i.e., the reaction time function that would have OLcurred in the real situation). If that is the case then cleariy we learn little abou; mental mechanisms (Type II processes) in such studies but only about what subjects know and what they understand the requirements of the task to be-i.e., about Type I processes, which require a ‘rational’ explanation. As I argued in 3ylyshyn (198 I), casting theories about such processes in the form of computer models does nothing to alleviate the confusion caused by the failure to distinguish these two types of processes. In that case one still needs to distinguish ?hose parts of the program that are merely emulating the functional architecture of the mind on the foreign architecture of currently available digital computers, from the Type I cog,titive processes that are hypothesized to be actually taking place in the mind. Ideally what we need to have is independent evidence for the basic cognitively impenetrable functions that are used to carry out the cognitive processes-i.e., we need to know the .functional architecture of the mind. This, as one might suspect, is no easy charge. Evidence for the nature of the functional architecture might come from behavior,al studies, from biology, or from ideas developed in corn-puter science (especially proposals, such as those of Newell, 1973? or of Fahlman, 1979, for radically different and psychologically promisiny a.rchitectures).
272
Zenon W.olyshyn
In the meantime, and in the absence of clearly adequate general architectural proposals, it is at least important to keep the distinction in mind. Sometimes knowing that there is an important distinction to be respected is just as important for the development of sound research programs as having reliable data. As an example of how keeping in mind the distinction between functional architecture and cognitive process can affect one’s theory construction, consider the following example drawn from some of our recent computer modelling work. We have taken some preliminary steps in the direction of developing a model of perceptual-motor coordination, a process which we believe contains some interesting examples of cognitively impenetrable functions -and one that may hold some clues concerning the often cited ‘spatial’ character of mental images and the rather special properties exhibited by images that are ‘projected’ onto perceived scenes. In designing this model (described in Fylyshyn, Elcock, Marmor and Sander, 1978) we were guided by the recognition that process models cannot avoid making some assumptions about the nature of the underlying mechanisms or architecture. In keeping with this concern we proceeded by what Marr ( 1976) has called the principle of minimum commitment: we have, whenever possible, aticnted the least powerful mechanism, compatible with known empirical constraints, that is capable of carrying out the task. This strategy allows the observed behavior to be characterized in terms of the fewest (or at least the weakest) assumptions about wl&h of the required functions are actually built into the organism, as opposed to being carried out by more general knowledge-dependent processes. This strategy, which is invoked only in the absence of detailed information about which functions are candidates for the functional architecture, is an instance of a quite general methodological principle: do not proliferate special cases when you can subsume instances under a general principle. It is like the policy adopted in the study of motivation, in which one at.tempts to characterize motives in terms of general cognitive pri-lciples in preference to attributing each apparently new motive to some newly discovered instinct. It is also equivalent to the principle of trying to fit a curve by an equation with the fewest free parameters. However successful these first steps may be (if nothing else they have suggested a number of empirical predictions which we have not yet tested), they do show that a fundamental distinction can have immediate consequences both for the interpretation of empirical findings and for the development of formal models. It should be added,.in concluding, that there are many cases in which we do have reason to believe that certain functions are cognitively impenetrable. For example there are many functions in the visual system (e.g., those which Marr, 1976, has characterized as deriving a ‘primal sketch’) and in language
Psychological exphmutions and knowledge-dependent processes
2; 3
processing (e.g., aspects of syntactic analysis or lexical lookup) that appear to be cognitively impenetrable. Because such processes i:a:ayprovide the ‘fixed points’ or the ‘cognitive constants’ in a future explanatory theory of mind, it is important to attempt to discover as many of them as ycssible. Indeed I anticipate important discoveries in the next decade to proceed along these lines: the strategy of ‘divide and conquer’ is an important one in science and I expect that careful analysis of task demands (along the lines being carried out in Artificial Intelligence) together with ingenious empirical work will reveal many modules of cognitively impenetrable processes like those cited above. At the same time, however, we must be cautious in accepting what seem on the surface to be primitive mechanisms of cognitive processing. Cognition is characterized by an amazing degree of plasticity, and the history of the field is littered with the remains of various proposed mechanisms that +vereinvented to serve as theoretical building blocks, but which turned out themselves to be just additional cognitive processes in need of a Type I explanation. In many of these cases, the apparent primitiveness of the processes, which led people to view them as elementary mechanisms, was an artifact of the experimental method used, and resulted from a blindness to the fwndamzntal distinctions discussed above and to the theoretical importance of the criterion of cognitive penetrability. The importance of these issues to the computational view of mind is discussed at length in Pylyshyn (fcrtl- Toming). Referemces Brewer. W. F. (1974) There is no convincing evidence for operant or classical conditioning in adult humans. In W. B. Weiner, and D. S. Palermo (eds.), Cognitiorcartdthe symbolic processes. Hillsdale, NJ, Prentice-Hall. Cooper, L. A., and &.eparcl, R. N. (1973) Chronometric studies of the rotation of mental images. In W. G. Chase (ea.), Visualinformation processing. New York, Academic Press. Fahlman, S. E. (1979) NETL: A system for representingand using reef world knowledge. Cambridge, Mass., MIT Press. Fodor, J. A,, and Pylyshyn, Z. W. (1981) How direct is visual perception: Somexeflectionson Gibson’s ‘Ecological Approach’. 4’>g., 9, 139-196. Kosslyn, S. M. (1980) image urrdmind. Cambridge, Mass., Harvard Unversity Press. Marr, D. (1976) Early processing of visual information. Phil. Trans. R. SIC. London, 275,483--534. Newell, A. (i973) Production systems: Models of control structures. In W. G. Chase (ed.), Visualinf;xmation prorn&ng. New York, Academic Press. Pylyshyn, Z. W. (1: .5) What the mind’s eye tells the mind’s brain: a critique of mental imagery. Pry chol. Bul., 80, l-24. Pylyshyn, Z. W. (1978) Imagery and artifical intelligenct. In W. Savage (ea.), Perceptionu;ld cogfiition: issuesin the foundations ofpsychology, Minneapolis, University cf Minnesota Press. Pylyshyn, Z. W. (197%) The rate of ‘mental rotation’ of images: a tesf c,,f a holistic analoguc hypothesis.Mem. Cog., 7, 19-28.
274
Zenon IS!Ifylyshyn
Pylyshyn, Z.‘W. (19793) Validattig computational models: a critique of Anderson’s indeterminacy of representation claim. Aychol. Rev., 86, 383-394. Pylyshyn, Z. W. (1980) Computation and cognition: issues in the foundations of cognitive science. Behav. Brain Sci,, 3,111,-169. Pylyshym, Z. W. (1981) The imagery debate: analogue media versus tacit knowledge. Psychol. Rev., 88, 16-45. Pylyshyn, Z. W. (In press) Computationand cognition Cambridge, Mass, MIT Press. Bradford Books. Pylyshy.n, Z. W., Elcock, E. W., Maaror, M., and Sander, P. (1978) Explorations in perceptual-motor spaces. Roceedings of the second internationalconferences of the Gmadian Society for Computational Studies of intelligence. Toronto, University of Toronto, Department of Computer Science. Shepard, R. N. (1964) Attention and the metrical structure of the similarity space. J. Math. Psycho& 1. 54-87. Swets, 1. A., Tanner, W. P., and Birdsall, T. G. (1961) Decision processes in perception. Psychol.Rev., 68. 30 l-340.
Cognition, 10 (1981) 275 -280 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
235
Psychology without linguistics = language without grammar DAN I. SLOBIN’ University of Californi-8, Berkeley
“Developmental psycholinguistics”, or the modern cdd;r of child language development, began about 20 years ago with a search for the early structurd principles underlying children’s speech in English. Cuided by Piagetian psycholog/ and Chomskyan linguistics, we approached language development as a phenomenon of successive grammatical structuration. The dominant theoretical issues of the sixtics dealt with the degree of innateness of the underlying principles, the extent to which they were specialized for language behavior, and their fit to current models of transformational grammar and its descendants. Psychologists and linguists discussed these problems in common terms , ar
solved by the child. With these two central figures explicitly ignoring the developmental dimension of linguistic competence, many social and behavioral Gent&s have been concerned with the web of nonlinguistic factors in which language acquisition is situated. These concerns have added richness and complexity to the tasks of developmental psycholinguistics, but they have hardly touched the problem of how it is that the child discovers and constructs Iinguistic form. The seven terms listed above reflect three intersecting groups of variables which lie outside of linguistic structure per se, but must play their roles in the acquisition of that structure: ( 1) 7 ;leoreticians of semantic and contextual factors emphasize that early messages MC supported by situational as we!! as linguistic irlformation in both comprehension and production. (2) Theoreticians of input,pragmatics,and discourse emphasize that much of meaning is carried by the structure and content of social interaction. (3) Theoreticians of cognition and language processing strategies argue that syntax is a reflection of more broadly-based structures and processes. There is no question that all of these factors are involved in language acquisition. However, taken in combination, advocates of these extralinguistic variables often seek to reduce language to something else, allowing it to arise “naturally” from processes not specially ad;rpted to the peculiar structures of ~yntax and morphology. However, tiiese structures and the course of their acquisition remain a puzzle to psycholinguists. The seven popular terms cannot be used to solve the puzzle without an eighth term, grammar. The formal medium that carries human discourse functions and is sequentially processed in perceptual and articulatory modalities- that is, structursd language-has peculiar attriba.ltes of its own. These attributes are best studied in conjunction with crosslinguistic comparison, where both diversity and. uniformity, particular and unversal features of language come to light. The existence of crosslinguistic diversity allows for a spectrum of “‘natural experiments” in which form and function can be profitably varied and juxtaposed. I have found comparative acquisition studies especially useful in developing notions of the processes underlying the path from child to adult grammar. (For details see references.) The importa-It preliminary conclusi,on from such studies,and related studies which look at the course of acquisition of individual languages in psycholinguistic terms, is that early phases of grammar can be accounted for in terms of particular factors of language processing, storage, and organizationinteracting with the three broad classes of extralinguistic factors listed above. Though we have much to learn, we are developing a set of expectations of
Psychology without linguistics = language without E;rarnnzar 277
what early child grammars should look like. It is significant, in addition, that such grammars have much in common with the grammars of creole languages, which also reveal underlying features of the language-creating capacities of the human mind. Furthermore, processes of change as revealed in ontogeny and in creolization are also reflected in the historical change of established languages. Common factors of these three diachronic pictures point to a characterization cjf the optimal core of human language-an idealization of the structures of semantics and code units and their interrelations. We know less about the processes by which elementary grammars are -evised. The study of reorganizations and restructu ngs in development promises to be significant in coming years. It is already evident that-to use Annette Karmiloff-Smith’s felicitous expression-“language is a forrplal problem-space per se for children” (in press). That is, beyond the goals of communication and expression, children work at mzkizig their speech conform in detail to the structural peculiarities of the language they hear about them. As a quick overview of progress and problems in understanding child language development, it may be useful to point out five major topics of research .:oncem: (1) development of the semantic svstem, (2) development of %.uface” units of linguistic expression, (3) mapy.ing between form and meaning, (4) going between form and meaning’ m real time (processing), and (5) revising linguistic systems to bring them into conformity with the norms of the speech community. (1) The semantic system is not isomorphic with the conceptual structure of thought, though the development of cognition and linguistic semantics are intimately related. Only a subset of conceivable concepts are candidates for grammaticization in human languages. For example, locative systems expressed by adpositions and affixes encode notions of topological location ( %I’, I ‘an’, ‘under’, ‘at’), source (‘frcm’), and goal (‘to’), but not contour of path, rate of movement, absolute size or distance. Nominal classifiers mark form but not color of referent objects. Many rn:>:*eexamples could be adduced; in short, the notions expressed by grammatical markers are a privileged subset of notions accessible to tile child. Within this subset, language-free developmental sequences are observed. For example, across a varrzty of distinct languages, topological notions of relative proximity (‘in’, ‘on’, ‘under’, ‘at’) are acquired earlier than notions of projective spatial location (‘in front’, ‘behind’) (Johnston. and Slobin, 197% The first past-tense markings always refer to processes resulting in immediately tangible end-states contemporaneous with the moment of speaking. The first encodings of transitive events- whether by word order, accusative markings on object nouns, OPergative markings on subject nouns-refer to
278
19.I. SIobin
direct physical manipulations of objects. In short, the development of grammatically relevant semantics lies outside of the development of language-specific means of expression (e.g. prepositions, postpositions, or nominal inflections for locative notions; affixes, stem alterations, or auxiliaries for ter- marking; and so forth). (2) At the same time, languages differ in the relative accessibility to the child of the means of expression of developing notions. For example, suffixes are more salient than prefixes, postpositions more salient than prepositions. Accordingly, the encoding of locative notions by suffixes (as in Hungarian) or by pos;positions(as in Turkish)emerges earlierthan the encoding of the same notions by prefixes (as in Bantu 1anguag:s) or by prepositions (as in English). However, the order of emergence of locative notions is the same across tlbese different types of languages. To some extent, form and function have separate developmental histories. As more semantic systems and more types of formal expression are studied ontogenetically, we will come to understand more about the child’s initial and growing sensitivities to particular features of linguistic form. (3) Languages also differ in terms of the ways in which semantic distinctions are grouped relative to linguistic form. Some conflations of semantic categories seem to be “primitive”, while others are not. For example, Turkish uses a single morpheme to encode accusative and definiteness on nouns, and this conflation poses no difficulty to children. However, conflations of accusative with animacy or with affirmative-negative distinction, as in Slavic languages, are not easily acquired. Or, to take another example of conflation, Slavic verb-stem alternations for tense are more difficult to master than verbstem alternations for aspect, indicating that aspect may be a more mherently verbal notion than tense. Numerous data of this sort will contribute to defining a developmental hierarchy of notions available for grammaticization. For example, case is likely to be grammaticized before animacy, aspect before tense, a.nd so forth. Eventually, comparisons of this sort will contribute to universal definitions of the semantic bases of grammatical distinctions. Such a hierarchy, however, also interacts with the specific grammatical form of encoding basic notions. For example, definiteness is acquired more easily if marked on the noun-phrase (as in Turkish noun suffixes and Indo-European and Semitic articles) than on the verb (as in Hungarian). Japanese children have difficulty in learning to adjust adjective stems for tense, indicating that tense may be an inherently verb-associated notion. Continued examination of crosslinguistic acquisition data on these lines will reveal “natural” tendencies to’ conflate or separate various semantic categories and to relate those categories to particular morphological and syntactic structures.
Psychology without linguistics = languagewithout grammar
279
(4) Systems of mapping between f-r-m and function must be processed “online” as speakers produce and interpret utterances. We are only beginning to elaborate mod.els of the developmert of proc . *,singskills. For example, word order patterns are relatively more difficult ‘:r attend to as guides to meaning than are particular “local cues” such as suf ..y3s. Turkish children of age 2; 0 are able to understand all six orders of subject, verb, and object is simple transitive sentences, where the object is marked with a distinctive accusative inflection, while English-speaking children of the same age cannot yet inter pret noun-verb-noun sequences as expressing subject-verb-object (Slobin, in press a). The development of processing skills clearly plays a signicicant role in determini1i.g which aspects of the grammar will be accessible to children at particular points in development, along with the perceptual and conceptual factors hinted at above. (5) Finally, linguistic systems continually undergo revision and reorganization throughout the course of development. Some revisions reflect attempts to secure clarity of expression, as when childr:n pass through stages of preferring analytic to synthetic expressions (e.g., make dead to kill), or full to contracted forms (e.g., will not to won ‘t). Other reorganizations reflect emerging realizations of regularity (as in the replacement of rote exceptions by rule-formed “errors”, such as the replacement of early correct rote past tenses broke and hit by breaked and hitted). Overregularizations are eventually replaced by a return of earlier rote forms, resulting in an adult-like system in which rules and exceptions coexist. Many separate examples such as these await the el&oration of a theory of stages of linguistic change. This quick overview of five major developmental issues points to a need for a comprehensive and interactionist developmental model in which particular linguistic skills of analysis and model building will take their place along with factors of cognitive development. It is far too early to make definitive state ments of the relative roles of language-specific and general cognitive factors in such a model. What is needed is many detailed studies of the course of acquisition of particular linguistic domains, iq all of their structural richnessmorphosyntactic, semantic, and-pragmatic. My prediction is that a powerful and definable “Language A6;quisition Device” (LAD) or “Language Acquisition System” (LAS) will eventually emerge. LAD/LAS, however, will l:ave to bootstrap itself in interaction with particular types of input--linguistic, cognitive, and social--modifying itself in predictable ways in constructing, and finally arriving at adult linguistic competence. Analysis of these psycholinguistic processes cannot succeed without detailed crosslinguistic developmental study, and cannot be carried out in ignorance of the detailed form of linguistic structures.
Asnmon, hf. S., and Slobin, D. L(l979)A cross-linguistic study of the processingof causative sentences. 43g., S, l-17. Johnston, J. R., and Slobin, D. 1. (1979) The development of Locative expressions in English, Italian, SerboCroatian and Turkish.I. child JZUM~., 6.531-547. Karmiloff-Smith, A. (In press). Language as a formal problem-space for children. In W. Deutsch (ed.), The chiW3 constmction of iknguuge. London, Academic Press. Skbin, D. 1.0973) Cognitive prerequisites for the development of grammar. In C. A. Ferguson and D. 1. Slobin teds.), Stu@iesof chiM hnguage deuelapment. New York, Holt, Rinehart & Winston, pp. 175-208. Skbh,D.l.r1l977)Lauguage changein childhood and inhistory. In J.Macnamara(ed.),LangguJearning and thought. New York, Academic Press, pp. 185-214. Sl&n, D. I. (1980) The repeated path between transparency and opacity in language. 1n.U. Bellugi and hf. !&adder@Kennedy (eds.), S&cd and spoken bguage: Biological constraintson linguistic fm West Berlin, Verlag Chemie, pp. 229-243. Slobin, D. I. (In press (I) Theorighrsofgrammatical encoding of events. In W. Deutsch (ed.), The child’s constmction ofkknguage. London, Academic Press. Sk&?, D. I. (In press b) Universaland particular in the acquisition of language. In L. R. Gleitman and E. Wanner (eds.), Language acquisition: Stute of the art. Cambridge,C- -bridge University Press. Slobin, D. I. &I.) (In preparation) The cmsdinguistic study of child lkngr:ap:. : I&dale, NJ, Lawrence Erlbaum Associates..
Cbgtirion, 10 (1981) 281-286 @ Elsevier Sequoia S.A., Lausanne - Printed in Tile Netherlands
281
Pragmatics DAN SPERBER C.N. R.S. and Universiti de Paris DEIRDRE WILSON* University College, London
Pragmatics, the theory of utterance-interpretation, is a branch of cognitive psychology. Utterances convey information which is conceptual, intentionally communicated and linguistically encoded, and which is processed in the context of additional conceptua! material retrieved or derived from memory. An adequate pragmatic theory she&id ‘ncorporate a general account of the processing of conceptual information $1 a context, and a ptrticular account of whatever special principles and problems are involved in the processing of information that has been intentionally, and linguistically, communicated. Pragmatic theories in this sense, if not under this name, have always existed; however, it is only in the last ten years or so that pragmatics has become an institutionalized research field, with its own textbooks, international conferences and journals.’ Its contributors are based in a variety of disciplines, including psychology and psycholinguistics, linguistics, AI and sociolinguistics. The field is so new and so diverse that no consensus on basic concepts and theories, or even on overall goals and research tasks, has yet emerged. Our remarks here like most contributions to the field, will be fairly speculative. The main aim of pragmatic theory is to provide an explicit account of how human beings interpret utterancei. To do this, one would have to say ihow disambiguation is achieved ; how reference is assigned; how sentence fragm.ents are interpreted; how ungrammatical utterances are dealt with. what role presuppositional phenomena play; how implicatures (intended inferences) are worked out; how contextual and encyclopaedic knowledge is brought to bear; and so on. Any organized’set of answers to these and similar questions would constitute a pragmatic theory on some level of adequacy. *Reprint requests should be sent to Deirdre Wilson, Department of Linguistics, Uriversity College, Cower Stxet, London WCIE 6 BT, England. a Recent and forthcoming general works include de Beaugrande and Dressier (1981); Cole (1978), Cole (1981), Leech (In press), Levinson (In press), Lyons (In press), Parrett and Verschueren (19801, and van Dijk (1977).
282
Dan Sperber and Deirdre WWn
Work so far published in the field tends to fall into three categories. The fust, and most interesting, consists of work which directly addresses these cetitral questions. Grice’s William James Lectures (1975, 1978) are classic examples. Here, Grice establishes a distinction between conventional meaning, assigned by semantic rule, and conversational meaning, created by the operation of a general communicative principle developed into various maxims of truthfulness, informativeness, relevance, perspicuity, etc. Work in this category can be both stimulating and suggestive ,2 but it is almost always so vague and intuitive as to constitute less a theory than a set of hints on how to go about constructing one. The second category consists of empirical work. A good example would be Clark and Schu,nk (1980), which investigates responses toI indirect requests and the properties which make them seem more oi: less polite. Work in this category, though it can be explicit and well-evidenced, is necessarily limited in scope, and is also hard to interpret in the absence of an established theoretical framework. The third category consists of formal work. An interesting example is Gazdaltlr(1979), in which the techniques of formal semantics are applied to a small range of pragmatic phenomena, and in particular pr,-,gmatic presuppositional phenomena. Work in this category is almost always explicit, but it is rarely directly relevant to the goals of pragmatic theory. Its practitioners tend to look only ai questions that seem immediately amenable to formal treatment, and these are rarely the fundamental ones.3 Over the last few years, we have tried to develop answers of our own to some of the central questions of pragmatics. Our work borrows from Grice’s the crucial insight that the interpretation of an utterance is based not only on me,aning and context but also on a general communicative principle tacitly shared by the interlocutors. Our work also differs from Grice’s in several important respects4 First, our claims are more explicit, For example, whereas G&e suggests a maxim of relevance without attempting to say what relevance is, we take an explicit definition of relevance as the basis for a reformulated Principle of Relevance, which in turn forms the basis for a unitied pragmatic theory (see bdow). Second, our work is psychological rather than philosophical in intent. We want to look at natural classes of phenomena, and to account for them in %ee, for example, Ducrot (19721, Stahaker (1974), and Allwood (1976). ‘Caz&u (1979, pp. 53-4) remarks that his attempted formalization of some aspects of Grids work loses much of the power and general&y’of Grice’s proposals, but adds ‘not to stick to formalist methodo@~in an areame this can only lead out of linguistics and into literary criticism’. ‘See Wilson and Sperber (la press) for discussion.
Pragmatics
283
terms of principles with systematic psychological correlates, This has led us to reject Grice’s most basic claim about the domain of pragmatic theory. clrice and most pragmaticians see pragmatics as concerned solely with utterancecomprehension, which involves recovery only of a set of propositions that the speaker specifically intended to convey. ’ We do not believe that comprehension is a well-defined domain. The typical case of communication seems to be one where the speaker specifically intends to express a certain proposition, and generally intendssome conclusions to be drawn from it (generally intends it to ha.ve some relevance); however, there is much variety in the further intentions that he could have. At one extreme, he may have no specitic intentiorls about the form or content of the conclusions to be drawn; at the other exLreme he may have highly specific intentions &out them; and between the two extremes, he could intend there to be conclusions of a certain general type, but not of a specific form, and so on. In other words, comprehension shades off imperceptibly into a wider process of utterance-interpretation, in which responsibility for a particular conclusion sometimes falls wholly on the speaker, sometimes falls wholly on the hearer, and in many cases is shared in some proportion by both. It is utterance-inn’crpretation, not utterance-comprehension, that is the natural domain of pragmatic theory. The third respect in which we differ from Grice is iu the role we assign to relevance in the processing of all conceptual informatiq and to a Principle of Relevance in the interpretation of utterances. Intuitively, to establish the relevance of some proposition is to see how it connects up with some accessible body of information (or context). We argue, more explicitly, that to establish the relevance of a proposition is to combine it with a context of acccs-e sible information and infer from this combination some conclusions (contextual impZicutionsPwhich would not be inferable from either the proposition or the context on its own. To maximize the relevance of a proposition is to process it in such a way as to maximize the number of its contextual implications and minimize the processing cost of deriving them. Maximizing relevance, in our terms, is simply a matter of extracting information from the combination of a proposition and a cor.text in the most efficient way, ancl it seems reasonable to assume that ali’ conceptual information is processed with this aim? Most information is not very relevant. However, when a speaker intentionally provides a hearer with information, he thereby gives a guarantee that a certain standard of relevance has been aimed at. This guarantee is incorpo‘Specifying the type of intentions involved is a complex technical matter. See Schiifer (1972) for discussion. For more general discussion of this issue, see Clark and Carlson (In press), and Sperber and Wilson (In press). 6For details of this account of relevance, see Wilson and Sperber (In preparation).
284
Dan Sperber andDeirdre Wilson
rated into our Principle of Relevance: the speaker tries to make his utterance as relevant as possible to the hearer. The hearer has a systematic expectation of relevance.. He may, of course, have more specific expectations as to what the speaker will say, but, we argue, either these further expectations follow from the Principle of Relevance in the context, or else they are just ordinary elements of the context with no special role in interpret,ttion. The Principle of Relevance plays a unique role in theminterpretation of utterances by.providing a shared criterion against which possible interpretations can be tested. An utterance has bet:n properly disambiguated, references have been properly assigned, sentence-fragments have been properly completed when and only when the resulting proposition satisfies the Principle of Relevance. We claim that either only one disambiguation of an utterance satisfies the Principle of Relevance, or an ambiguity is perceived by the hearer.’ Similarly, the Principle of Relevance determines the implicatures of an utterance: when the speaker could not have expected his utterance to be relevant to the hearer without intending him to derive some specific contextual implication from it, then, and only then, that implication is also an implicature.’ A central question concerning the processing of conceptual information in context is how contextual information is selected, retrieved and expl,oited. We argue that this complex process is governed by the goal of maximizing relevance. A small initial context (the interpretation of the preceding utterance in the case of verbal material) is systematically searched for contextual implications and can be expanded in several directions, While each expansion may increase the number of contextual implications, it also increases the processing cost in such a way that the context must be kept narrow or else relevance would decrease.9 When, in the inferential processing of a proposition, the systematic search of a narrow context for possible contextual implications fails to satisfy expectations of relevance, use is made of what we call ‘evocational processing’. This second form of processing consists in sampling a much larger context in search of conceptual connections on the basis of which relevance might be increased. Evocational processing can be intentionally triggered in a subject by providing him with information the relevance of which he will not be able to establish solely through inferential processing. This occurs, in particular, when-7 Expeszmental studies on’the
processof disambguationtake the goal of that process for granted. The goal of disambiguation is intuitively obvious but has not been so far explicitly described. The Principle of Relevance provides the basis for such a description._ *Presuppositional phenomena can also be accounted for on the basis of the hinciple of Relevance; see VGlson *nd Sperber (1979), Sperber and Wilson (In preparation). For a discussion of previous accounts, see Wilson (1975). 9!%eSperber and Wilson (In press).
B3lagmatics
285
ever a figure of rhetoric is used. Figures of rhetoric such as metaphors cre?te and, at first, frustrate expectations of relevance which can then be satisfied through evocational processing. Along these lines, the Principle of Relevance can provide some insight into not only th3 inferential but also the evocatioilal aspects of utterance-interpretation, and hence folrm the basis for a unified theory of pragmatics and rhetoriclO If our approach to pragmatics is right, it closes dowu one line of research currently being pursued, and opens up a quite different one. Much recent work on pragmatics assumes the existence of a pragmatic device or module. with its own formal properties and rules, comparable to the linguistir device and others suggested by recent psychological research. As far as we can see, there is no evidence for this assumption, or for the widely held alternative assumption that pragmatics is simply one component of the linguistic device. Prag,matics is not a separate device or sub-device with its own specialized structure: it is simply the domain in which linguistic abilities, logical abilities and memory interact. Precisely because of this lack of specialization, we think pragj atics can yield valuable insights into other areas of psychology. There is a whole range of highly complex natural phenomena which are in important respects beyond the scope of experimental methods, and about which informants -make only vague and subjective statements: for example, the unders!anc”;ng of a work of art, or a ritual, or another person. Utterance-interpretation i: also a highly complex natural phenomenon which cannot always be er.perimenta’!y studied; however, intuitions about utterance-interpretation are somewhat more clearcut and less controversial: it would be rare, for example, to find two informants disagreeing about the disambiguation of an utterance in context. Psychologically justified pragmatic theories are thus easier to construct than, say, psychologically justified aesthetic theories. If we are right about the lack of a specialized device for utterance-interpretation, the basic principles involved in it should be equally applicable to other complex natural phenomena of the same type: phenomena which, like the interpretation of utterances, seem to involve, in a search for relevance, a combination of inferential and evocational processing. Pragmatics thus seems to us to be capable of throwing direct light on psychological mechanisms of some generality and indirect light on other areas of thought where these mechanisms play a part. We think pragmatics is entering a more active and creative phase. However the best we can expect is quite modest compared to the task at hand. It is likely that theoretical work in pragmatics (and in related areas of psychology) will remair highly speculative. But speculation need not be trivial or vague. ii-
See Spurber (197Sa,19756,1980),
Sperber and Wilson (1981, In preparation).
286
Dan Sperher and Deirdre Wilson
Allwood, J. (1976) Linguistic Communication as Action and C&operation: A ::tudy in Pragmatics Gothenburg Monogmphs in Linguistics2. University of G6tebolg. de Beaugrande, R., and Dressier, W. (1981) Zntroductionto Text Linguistics. London, Longman. Clark, H., and Carlson, T. (In Press) Context for comprehension. In J. Long and A. Baddeley teds.), A ntwtzbn and Performance, XI. Hillsdale, NJ, Erlbaem. Clark, H., and &hunk, D. (1980) Polite responses to polite requests. Cog., 8, 11 l-143. Cole, P. (cd.) (1978) Syntax and Semantics 9: Ragmatics.New York, Academic .‘ress. Cole, P, fed_) (1981) RadicalPragnraticsNew Ywk, Academic Press. Ducrot. 0. (1972) Tie et ne pas d&e. Paris, Hermann. Gazdar, G. (1979) Rag?rurtis: Zmplicature.Presuppositionand Logical Form. New York, Academic Press. Grizze, H. P. (1975) Logic and conversation. In P. Cole and J. Morgan teds.), Syntax and Semantics3: Speech Acta New York, Academic Press. G&e, H, P. (197% Further notes on logic and conversation. In P. Cole ted.), Syntax and Semantics 9: Ragmatics~New York, Academic Press. Leech, G. (In press) Linguisticsand Rhetoric. London, Long&n. Levinson, 6 (In press) ZVagmatics.Cambridge. Cambridge University Press. Lyons, J. (In press) Lorrgrurge,Meaningrmd Context. London, Fontana. Parret, H., and Verschueren, J. (198O)P~~gmuticsond Beyond. Amsterdam, 1. Benjamin. Schiffer, 8. ( 1972) Meumtig. Oxford, Clarendon Press. Sperber, D. (19750) Rethinking Syn&ofism. Cambridge University Press. Sperber, D. 1(1975&jRudiments de rhitorique cognitive. Po&que, revue de thiorie et d tinalyselit& robes. 23,389-415. Sperber, D. (1980) Is symbolic thought prerational? In M. Foster and S. Brandes (eds.11,Symbol as Sense. New York, Academic P:ess. Sperber, D., and Wilson, D. (1981) Irony and the use-mention distinction. In P. Cole ted.), Syntax and Wnant&s 9: fiagmutics. New York, r;&emic Press. Sperber. D., and Wilson, D. (In press) Mutual knowledge and relevance in theories of comprehension. In N. V. Smith Bed.!,iVow&ngs of the SSRC Colloquiumon MutualKnowledge. London, Academic Press, Sperkr, D., and Wilson, D. (In preparation) hnguage and Relevance: Foundations of Pragmatic *ovS~htier, R. (1974) Pragmatic presuppositions. In M. Munitz and P. Unger teds.), Zemanticsand Phiiosuphy. New York, New York University Press. KiiIn, D. 11975) Resuppositions and Won-l?uth-GmditionalSemantics. New Yorlk, Academic Press. Rilson, D., and Sperber, D. (1979) Ordered entailments: an alternative to presuppo$tional theories. In C.-K. Oh and D. Dineen @W,Syntax and Semantics ZZ: Aesupposition. New York, Academic Pzess. B=n, I)., and Sperber, D. (In prew) On G&e’s theory of convezation. To appear in P. Werth (ed.), London, Croom Helm. ~~mmtztkm~&wech and hccmrse whn, D., and Sperber, D. (In preparation) On defining ‘relevance’. To appear in R. Grandy (ed.), Femchri@for PbuI &ice. Van 0% T. (19773 Text and Context.: Expkxations in the Semantics and : .-agmaticsof Discourse. Loruion, Lorrgman.
Cbgnidon, 10 (1981) 287-294 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
287
Cognition in animals: Learn +j as program assembly J. E. R. STADDON” Duke University
Learning most tasks has two components. First, an approximate solution must be found: the student laboriously solves the first algebra problem, the rat stumbles on the lever. Second, after an intervening period when the organism does something else, a residue must remain of what occurred on first exposure so that there are savings, i.e., the organism does a bit better the seccmd time than it did the first time. Learning is compounded of these two effects: initial often ill-directed stabs at solving the problem, which will be repeated in reduced form on subsequent exposures, plus the retention from occasion to occasion of an accumulating core of reliable knowledge. The two major areas of animal learning are divided along these general lines. Operant conditioning, to the extent that it is concerned with learning at all, is interested in the initial “shaping” of behavior by the condition: ,f reinforcement. There is much less interest in how the changes wrought in one experimental session carry over to the next. Classical conditioning, on the other hand, is primarily concerned with the accumulation of “strength” by a conditioned stinmlus from one experimental session to the next. Although classical and operant conditioners disagree on many points they agree in one respect: on the importance of stimuli and responses. Both lines ofwork have their theoretical roots in the refle.u---not the reflex of Sherrington a subtle concept subordinate to his main theme of integration, but a simpler idea used as a metaphor for cause. In recent years, dissatisfaction with this primitive kind of theory has led to the appearance of “cognitive” views. These take several forms: updated versions of Tolmanian cognitive maps (e.g., Menzel, 1978), experimental exploration of the perceptual categories of animals (e.g., Herrnstein and de Villiers, 1980), more or 1~s~ explicit informationprocessing accounts (e.g.,Wagner, 1978),through accounts of animal behavior as a rational process paralleiing human verbal reasoning (e.g., Seligman and Johnston, 1973). Any account involving the word “memory” or referrip: - to constructs at all’removed from measurable stimuli and responses tends also to label itself “cognitive” (cf., several chapters in the book edited by Hulse, Fowler and Honig, 1978. *I thank Richard Herrnstein, Stewart Hulse and Evalyn Segal for comments on an earlier version of this piece. Research supported by grants from the National Science Foundation to Duke University. Reprint requests should be sent to J. E. R. Staddon, Department of Psychology, Duke Universtiy, Durham, North Carolina 27706, U. S. A.
288
J. E. R. Sraddon
In some respects, these changes are a clear advance. The evidence that animals can form representations of the world of rich, maplike connectivity is overwhelming. The idea that ability to time can be described in terms of an internal clock is perfectly plausible, even though its physiology (in mammals) and properties are far from fully understood. It is useful to emphasise that theories of animal behavior need not be restricted entirely to peripheral processes: We are no longer confined by the bizarre presumption that the only acceptable kind of biological clock is one in which the hands and escapement are composed of the animal’s observable behavior, for example. Yet this new freedom is not without its dangers. In the exhilaration of throwing over what I have termed “hooks-and-eyes” behaviorism (Staddon, 1973), a distressing tendency to revert to old-fashioned anthropomorphism has Surfaced, Animals are said to learn that they are “helpless” and cannot control external events; they learn “relations”, they “remember”, they “expect” certain things to happen, and so on. “So what! “, the reader may object. It is well known that computer programmers speak of their machines in just this familiar way: the machine remembers past data; it expects certain kinds of input and behaves disagreeably if it fails to receive them; it may even act “helpless” under certain circumstances. Yet no one doubts that a proper scientific account can be given of all these things. Books on programming are free of the exhortations to terminological hygiene that were once commonplace i.ntexts on behavioral psychology. The difference between the computer and the animal, of course, is that a mechanistic account of computer behavior is available to anyone with the wit and patience to thread his way through the appropriate program. AI1 know that such an account exists, even if few choose to check it out. But mechanistic accounts are lacking in many of the examples I have given. We have no more insight into the behavior of a “helpless” person or animal than is provided by our personal intuition -an attribute in which many individuals are notoriously deficient. Little agreement can be expected on this basis. The problem with “relational” learning is at least as old as the controversy between Kenneth Spence (1937) and the gestalt psychologists. “Animals learn relations”, they said -referring to the results of transposition experiments. “Then why do they sometimes fail?” asked Spence-referring to transposition reversal. Spence’s solution was an additive process by which particular physical stimuli are associated with gradients of effect that summate to guide choice. His theory is simple, perhaps too simple, but at least it provides a mechanism not only for those instances where animals behave relationally, but also for the cases where they do not. One may disagree with Spence’s solution, but one cannot avoid the obligation to do as he did and
Cognition in animals: Learning as program assembly
289
provide a mechanism for those “cognitive” aspects of performance for which anthropomorphic labels are so seductive, Sympathetic understanding is not scientific explanation. The modest insig.at that animals learn more than stimulus-response relations does not justify the leap into a never-never land of mental life. Not only do rationalistic explan&ions beg the question, they may also miss the point. Many years ago, the comfortable anthropomorphising of Romanes and his successors was demolished by careful experimental analyses that dissected such elegant adaptations as predation and nest building into simple programs, whose elements could be laid bare by breaking invariable natural associations. For example, Fabre ( 19 15) showed that if the paralyzed prey of a cricket-hunting wasp (left briefly by the predator as she inspects h.er burrow) is slightly moved, the emerging wasp insists each time on repositioning the cricket and inspecting the burrow again. This sequence is repeated indefinitely, as often as the cricket is moved, Even vertebrates show similar rigidities-gulls retrieving imaginary eggs and incubating baseballs, robins aHacking distant mail vans, fish courting colored discs, chicks imprinting on almost anything, and so on. The point is that “rational” behavior may rest on quite a simple set of well-knit subroutines. Rationality vanishes as soon as the rules by which the animal operates are known. The danger of rationalistic accounts is that they lead us to neglect the task of challenging the organism in ways that will reveal the mechanical structure that must underly the appearance of rationality. We learned this lesson once; must we do so again? If not expectations of S-R connections, then what do organisms learn? I suspect that the very form of the statement is misleading: the organism (an agent) does something (it learns). Net so; what actually happens is tl;st we (the experimenter) expose this biological system to an interactive environment of some sort and, after a while, the system behaves in a different way than it did before. And the changes persist, so that even after lapse of time, the new pattern of behavior is retained. In the metaphor of the computer, the organism behaves as if it has a new program. There is no agent here, unless it be tllc experimenter. No thing is learned, Rather from ingredients given by nature and past experience a new program better suited to changed conditions has evolved It may be convenient to describe the properties oF this new program by saying that the animal has developed an expectation or is “helpless”, but this no more explains what is happening than it would were the program a real one. written for an electronic, rather than a biological, computer. The two stages of learning that I described earlier fit easily into the program metaphor. The first stage, in which behavior is variable and “exploratory” may represent the first steps in putting together an effective program.
290
1. E. R. Staddon
Such a program has two aspects. One is concerned with knowledge: the program must encode the new situation in a way that relates it to--and differentiates it from -others in the animal’s experience. This code also functions as a retrieval cue for the second aspect of the program: its performance rules; these guide behavior with respect to the encoded representation. If the first: step has been well done, then situations functionally similar to the training situation wi!l be sufficient to retrieve part, at least, of the original program-the animal is then said to generalize in an adaptive way. For the most part functiona; and physical similarity are closely related: in nature, physically similar situations will usually require similar behavior. Occasionally, as in a discrimination-learning experiment, physically similar situations will require quite different behavior. Even though su,;h similar situations will at first retrieve similar programs, if the similarity i.s not too great, the processes of program assembly (i.e., learning) are usually sufficient tc rllow the development of different programs for each situation. The animal is then said to discriminate appropriately. This analogy has several implications for our understanding of conventional operant-conditioning procedures: I. Stimulus control
Rather than being a relation analogous to 2 reflex, this is really close,. to a memory retrieval process. The nature of the process is obscured because +he stimulus responded to-the colored pigeon key, for example-is also a large part of the situational context that calls up the behavior of key pecking. The resembkmce is clearer in the case of temporal control, where the stimulus in questial ti in the past (the pause in responding produced by periodic food on a fixed-interval schedule is an instance of temporal control). Here limitations on temporal control correspond closely to limitations on memory revealed in experiments with people (cf., Staddon, 1974). Experiments on successive discrimination reversal can also be ana?.yzed in a way tihat makes clear the resemblance between stimulus control and memory-retrieval1 (Staddcn and Frank, 1974; Staddon, in preparation, Chapter 11). 2. Acquisition The Hghly variable, exploratory behavior that characterizes first a new situation may represent a relatively fixed program that whenever the animal confronts a situation not previously encoded it has no stored prmm. (This view is close to Craik’s [ 19431
exposure to is called up or for which w&known
Gq-nitionin animals:Learningas programassembly
29 1
interpretation of learning as the detection of a discrepancy between a stored model and a new situation). Proof of this conjecture is difficult because few situations are truly novel (in the sense of having no resemblance to any previously experienced situation), so that animals’ response to most new situations usually contains some previously learned elements. Nevertheless, in some careful observations of rats in a runway in my laboratory a few years ago, Kello (1973) recorded a stereotyped sequence of exploratory behaviors that could be reliably elici.ted by novel odors. It may be that sufficiently detailed observations of other animals exposed to a range of situations will reveal similar stereotypy. This “new-situation” program presumably has the function of assembling, by a process analogous to variation and selection (Campbell, 1960; Staddon and Simmelhag, 1971) elements of a new program, well adapted to the new circumstances. 3. 1ndivi:lual differeqces Two programmers will rarely accomplish the same task in the same way: if the job is at all complex, many programs, of roughly equal speed and efficiency, can usually be devised to accomplish it. It is no surprise, therefore, to find that beneath the surface similarity of performances on standard reinforcement schedules there is quite a range of individual differences. For example, Figure 1 shows the cumulative response records during an extinction (no reinforcement) test of two pigeons identically trained on a cyclic variableinterval schedule (Innis and Staddon, 1970). Both pigeons showed the same tmcking behavior in training, respondling fastest at times of high food density; yet in the extinction test, one animal (#437) shows a reliable periodicity, the other does not. Evidently each animal accomplished the tracking in a different way. Similar results have been reported in experiments on attention (i.e., selective stimulus control) in pigeons (e.g., Reynolds, 1961). Most students of both human and animal learning when pressed can come up with other examples. Since they make poor material for the hoped-for general laws of learning, (and are statistically inconvenient besides, relatively few find their way into print. 4. Reversibility and observability Learning is not a reversible process; moreover5 the program analogy implies that different organisms are likely to learn different things (i.e., evolve different programs) even in the same situation. These two features pose severe
292
J, E. R. Staddon
methodological problems for the study of learning. Because of individual differences, group-average data (methodologically required by irreversibility) are likely to be of only limited usefulness. Yet theorems well known to students of cognitive science (e.g., Moore, 1956;Kalman, 1968;see also Houston, Halliday and McFarland, 1977; Staddon, in preparation, Chapter 8) place severe &nits on what can be discovered about a “black-box” system, even when an indefinite number of identical copies is available. When replicas are not available, the problem of discovering the properties of a complex black box are substantial indeed. There is obviously a pressing need to develop novel methods for studying learning in individual organisms. 5. Reinforcement and Optimality analysis
Organisms may generate the same performance in different ways; nevertheless, the performance itself often makes sense in terms of its outcome. Animals in situations not too different from their natural environment will often behave in ways that maximize net energy intake or minimize time spent, for example (cf., Maynard Smith, 1978; Staddon, 1980). Approaching operant behavior from the po; .. of maximization also derives in a natural way from the familiar notion of rbn,tircement as “control”of behavior by its consequences-indeed, Rachhn (1980) has persuasively argued that reinforcement theory and the Figure 1. Cknulative records fbr two pigeons during two separate 20-m&ztest periods in which no reinforcement wasavailable,following trainingon a cyclic-interval schedule. figeon #37 shows cyclic variation, pigeon $103 does not. (From Innis and Staddon, 1971, Figure 5).
Cognition in animals: Learning as progmm assembly
293
econotnics of utility are one and the same. Analysis of behaviors as constrained optimization is a useful way to deal with what von Bertalanffy (1956) has termed “equifinal” systems, that is, systems that can achieve the same objective through a variety of means. Optimality analyses can provide simple accounts for complex adaptive behavior. In addition, the nature of the constraints that must be assumed to obtain accurate predictions tells us something about the mechanisms of learning -the principles of program assembly, in terms of my metaphor. Recent interest in economic and optimality analyses of operant behavior (e.g., Herrnstein and Vaughan, 1980; Houston and McNamara, 198 1; Rachlin and Burkhard, 1978; Rachlin, Green, Kage:i and Battalio, 1976; Staddon, 1979; Staddon, Hinson and Kram, 1981) is a belated recognition of the equifmal nature of operant behavior. References Campbell, D. T. (1960) Blind variation and selective re .ention in creative thought as in other knowledge processes.Psychol. Rev., 67, 380-400. Craik, K. J. W. (1943) The rsafureof explanation. Cambridge, University Press. Herrnstein, R. J. and de Villiers, P. A. (1980) Fish as natural categories for people and pigeons. In G. H. Bower(ed.), The psychology of learrdngand I ?tivation: Advances in research and theory. Vol. 14. New York, Academic Press. Herrnstein, R. J. and Vaughan, W. (1980) Melioration and behavioral allocation. In J. E. R. Staddon, (ed.), Limits to action: The allocationof indivtiual beh jior. New York, Academic Press. Fabre, J. H. (1915) 13re hunting wasps.New York, Dodd, Muad & Co. Houston, A. I., HaIliday, T. R., and McFarIand,, D. J. (1977) Towards a model of the courtsh@ of the smooth newt ;Ttiturusvulgar&,with special emphasis on problems of observability in the simulation of behaviour. Med. BW Eng. IS, 49-61. Houston, A. I. and McNamara, J. (1981) How ito maximize reward rate on two variable-interval paradigms. J. Ex.rer. Anal. Behav., 35.367-396. Hulse, S. H., Fowler, bI., and Honig, W. K. (Eds.) (1978) C&nitir~eprocesses in animal behavior. Hillsdale, NJ, Lawrtnze Erlbaum. Innis, Nancy K., and S: addon, J. E. R. (1971) Temporal tracking on cyclic-interval reinforcement schedules. J. Exper. A,qai.Behav. 16, 4 11-d’ 23. Kalman, R. E. (1968) New Developments in system theory relevant tobiology. InM. D. Mesarovic (ed.), @stems theov md biology. New York, Springer-Verlag. KeIlo, J. (1973) Observation of.the behavior of rats running to reward and nonreward in an al&way. Ph. D. Dissertation, Duke University. Maynard Smith, J. (1978) Optimization theory in evolution. An. Rev. Ed. 9, 31-56. Menzel, E. W. (1978) Cognitive mapping in chhmpanzees. In S. H. Hulse, H. Fowler, and W. K. Ho& (eds.), Cbgnitiveprocessesin antmalbehavior. Hillsdale, NJ. Lawrence Erlbaum. Moore, E. F. (1956) Gedankenexperiments on sequential machines. In C. E. Shannon and J. McCarthy (eds.), Automata studies. Princeton, NJ, Princeton University Press. Rachlin, H. (1980)~Economics and behavIora psychology. In Staddon, J. E. EL (ed.), Limits to action: 17rea&catiGn of individualbehavior. New York, Academic Press.
Rachlin, H., and Burkhard, B. (1978) The temporal triangle: Response substitution in instrumental conditioning. Psychal. Rev. 85, 22-48. Rachlin, H., Green, L., Kagel, J. H., and Battalio, R. C. (1976) Economic demand theory and psychological studies of choicq. In G. Bower (ed.), The ,asychologyoflearningand mofivation(Vol. 10). New York, Academic Press. Reym&, G. S. (1951) Attention in the pigeon. J. exper. Anal. B&v., 4,203 -208. Seligman, hi. E. P. and Johnston, J. C. (1973) A cognitive theory of avoidance learning. in F. J. McGuigan and D. B. Lumsden (eds.), Contemporaryapproachesto conditioning and learning. New’ York, Wiley. Spence, K. R. (1937) The differential response in animals to stimuU varying in a single dimension. Pq cho!. I&w. 44,435-444. Staddnn, J. E. R. (1973) On the ?otion of cause, with applications to behaviorism. Behav. I, 25-63. Staddon, 3. E. R. (1974) Temporal control, attention and memory. Psychol. Rev. 81, 375-391. Staddon. J. E. R. (1979) Operant behavior as adaptation to constraint. J. exper. Pqychol. Gen. ZOS, 48-6 7. Saddon, J. E. R. (1980) Optimality analyses of operant behavior and their relation to optimal foraging. In i. E. R. Staddon, (Ed.), Limits to action: The allocationof individualbehavior. New York, Academic Press. Staddon, J. E. R., and Frank, Janice. (1974) Mechanisms of discrimination reversal. An. Behav. 22, 802-828. Staddon, J.,E. R.,Hirjon, J. M., and Kram, R. (1981) 0ptimalchoice.J. exper. Anal Behav..35,397-412. Staddon, J. E. R., and Simmelhag, V. (1971) The “superstition” experiment: A reexamination of its implications for the principles of adaptive behavior. Psychol. Rev. 78, 3-43. van Bertalanffy, L. (1956) General systems theory. GeneralSystems Handbook, 1. Wagner, A. R. (1978) Expectanci&-s and the priming of STM. In S. H. H&e, H. Fowler, and W. K. . Honig (eds.), Cognitivepnxesses in animal behavior. Hillsdale, NJ, Lawrence Erlbaum. Staddon, J. E. R. Adoptive behmtiorand &aming. New York: Oxford University Ress, in preparation.
Cognition, 10 (1981) 295-300 @ Elsevier Sequoia S.A., Iausanne - Printed in The Netherlands
295
Cognitive development in school and out SIDNEY STRAUSS” Tel Aviv
University
The contributors to this issue were given the charge to reflect on their own work and to suggest where main outcomer and theoretical breakthroughs in the area of cognitionawill be coming in the next decade. I have broadly taken this to mean discussing where my work will be heading in the next number of years. In my own egocentric way what I will be writing here has to do with orie direction I think my area ought to go; in the process I will reflect on problems that have arisen in my work that I think have soms generality. Given space limitations, I will only be able to present here: (1) 2 single question that strikes me as important for my field, (2) a brief anaJysis of various approaches that can be used to inform it, and (3) an ever. tziefer statement about a general problem besetting the study of cognition. Before getting on with the above, I must state at the out:+ that my field is what might be called applied cognitive developrmental psychology. The area in which my cognitive developniental work is applied is educarion and my concerns and the problems I investigate are always chosen with an eye towards their importance for educ&ional theory and practice. Before the reader deci.des to turn to the next article, I would like to note that although there are vzry few people who would define their field in this way, and despite the fact that it sounds like a very narrow intersect of several areas with-, out an essence of its own, it is my conviction that the area I am about to describe strikes at ihe deepest roots of the ways we represent what it is that we know and the ways that these representations can conflict with each other -and in so doing lead to cognitive development. The placement of these representations in the spectrum of epistemological issues related to common sense knowledge through knowledge that has a cultural history, as in the case of scientific and mathematical. knowledge, is a central goal of my work. It has intrigued me that we seem to represent our knowledge in different languages or notations (for lack of better words) and that these notations seem to influence how we think about various problems. One of the ques*I wouldlike to thank David Feldman for his comments on an earlier draft of this paper. This paper is Working Paper Number 5 of the Tel-Aviv University Study Group of EIumanDevelopment. Reprint requests should be sent to Sidney Strauss, School of Education, Tel-Aviv University, Tel-Aviv(Ramat Aviv), Israel 69978.
296
S. Strauss
tions that I have been asking is what influence one mode of representation has on another. One system of representation is that of common sense, everyday.notions of events in the environment that have regularity. This regularity is a mental construction, of course. An indication of how this kind of mental representation works comes from the ways one solves physical concept tasks that measure intensive physical quantity. Intensivity is a quantity that is not dependent on its amount and, temperature is an instance of this concept. A task that measures children’s understandings of this aspect of temperature, when presented qualitatively, is as follows: Three cups (A, B, and C) of equal amounts of sametemperature water are presented to children. All of the children judge them to be the same temperature. The water from two of the cups (A + B) are then poured into an empty container (D) and the children are asked to compare the temperatures of the water in the mixed cup (D) and the untouched cup (C). When tasks such as this one were presented to children ages 3%. 12% it was found that very young children solved this task correctly and justified their correct judgments as follows: ‘They were the same before so they are the same now’ (identity justification). Older children produced an incorrect judgment for this task and justified it in te:ms of extensive, additive quantity; i.e., they argued that the cup with more water is hotter or colder depending on whether or not the original water was hot or cold. Still older children produced the correct judgment and it was accompanied by the identity justification (Strauss, in press; Strauss, Stavy, and Orpaz, 1977). Another way to present this same concept of intensive physical quantity is via a numerical representation. For example, we can present children with water in two containers and measure its temperature with a thermometer. Let’s say the temperature recorded in each cup was 10 “C. The water is then poured into a third, empty cup and the children are asked to tell us the temperature of the mixed water. The overwhelming majority of children through age 13 answer that ihe mixed water’s temperature was 20 “C instead of the correct answer of 10 “C (Strauss et al., ? 977). Now we set that there is a conflict between children’s qualitative, common sense representations of temperature and their representation of the number system as they apply it to temperature. In the case of the representation in the number system, what we are probably witness to here is a confusion on the part of children between joining and additivity (Carnap, 1966; Cohen and Nagel, 1934; Hempel, 1952). Notice that in the case of intensivity water is being physically joined or combined. Hen@ (1952) introduced a symbol for physically joining or combining, a small circle, instead of the plus sign. Hence, xoy designates, in our case, mixing water from one cup with water from another cup. In the case of exten-
Cognitive development in school and out
297
sive quantity the arithmetic analogue is ‘+‘. The general principle of additivity, with respect to any magnitude, M can be expressed: M(aob) = M(a) + M(b) The reason for this distinction is that there are two types of additivity : (1) the physical operation of joining objects and (2) the arithmetic operation performed on numbers. One does not add, say, lengths or weights. Rather, one adds numbers that represent lengths of lines or weights of objects. Arithmetic operations of additivity can be appropriately performed on numbers that represent measures of extensive physical quantities. In other words, arithmetic operations of addition model the physical joining of extensive physical quantities. It may be the case that children attempted to apply this model to numerical intensive physical quantity tasks, thus producing kncorrect judgments on them. I shall now very briefly attempt to characterize the two kinds of knowledge just described. The common sense representation of qualitative empirical regularities is tied to complex interactions lbetween the sensory system, the environment that supplies the information to our sensory system, and the mental structures through which we organize the sensory &formation and which guides our behaviors. I ar,gue here that individuals’ common sense knowledge a.bout qualitative physical concepts is no different today than in the times of, say, Aristotle. These are spon.taneous concepts (Vygotsky, 1962) that are universal acquisitions (Feldman, 1980). In contrast, the numerical representation of the same (or other) concepts ie QC?!spontaneous but, rather, requires instruction before it is part of the conceptual repertoire of the individual. And it is cultural/historical in the sense that, to keep our examples parallel, the scientific and mathematical knowledge about physical concepts such as temperature have changed radically since Aristotle’s time. Now t!z~, what happens when these two representations are in conflict with each other? For example, in the case of temperature we saw that somewhat older children correctly solve the intensivity task when it is presented qualitatively and incorrectly solve it when presented numerically. Children from ages 7 through 1 l., were confronted with their own contradiction and were asked what they thought about it (Strauss it al., 1977). One sign (although not a decisive one) that children’s qualitative and numerical representations oi‘ temperature are different was that many of the children at all ages found it difficult to understand that the qualitative and numerical versions were the same task asked in different ways. ,4fter some discussion, children began to understand that these tasks were the same. We found three types of responses to the conflict, these types being age-related. The frost type was found among the youngest children who
did not even recognize that a conflict existed and ar:gued that cold water when mixed with cold water remains cold and that the same water when measured to be 10 “c and mixed turns out to be 20 “C. Or as many children stated, ‘It’s different when you have numbers’. The second type of response was found among somewhat older children and it was to change their correct common sense qualitative understanding to an incorrect one by arguing that cold water wheq mixed with water at the same temperature becomes colder. I-Iere we see a drop in performance that results from one rep.resentational system ovel?iding another representatianal system whose notation or language is different. The third response to the conflict was found &mong the oldest children who changed their incorrect numerical answers to correct ones. I have argued elsewhere (Strauss and Stavy, in press) that in the case of the oldest children, the correct solution on the numerical task is arrived at through ‘consultation’ with the correct qualitative understanding of the problem. Now we see the very interesting developmental phenomenon that the numerical representation that overrode the performance in the qualitative representation becomes reorganized by that very qualitative representation. Other examples of this phenomenon come from Sheeran’s (1973) work on children’s . ?nsorimotor and verbal representations of weight conservation and from Bamoerger’s (in press) work on children’s sensory (figural) and musical (formal) representations of rhythm. A curriculum unit on the concept of temperature was developed for fifth grade children based on some 0’ the above considerations. This unit was introduced into Israeli classrooms and was tested against a graup of children who were given individualized instruction and a control group (Stavy and Berkovitz, 1980). The results were that those who learned from the curri-, culum unit and those who received individualized instruction advanced considerably in their ability to correctly solve numerical intensivity tasks. The developmental information allows us to make informed decisions about curricular content, its sequencing, and timing. It is my conviction that a curriculum unit, if not based on a thorough understanding of the development of children’s competerices about the concepts to be taught, is essentially blind and should not *be expected to be successful as a teaching instrument. Notice that the concepts taught in th.is unit were those that I had previously described as being nonspontaneous and cuhural/historic;ti in character and ia need of instrrlction if they are to become part of how children come to represent their world. In other woxds, they are nonuniversal (Feldman, 1980). In one aspect of education one tries to pass on the cultural/historical (in my example, scientific and mathematical) knowledge while taking into account the cumrnon sense knowledge that children have developed. Sometimes the scientific and mathematical knowledge go hand in hand. Sometimes the com-
Cognitive development irrschool and out
299
mon sense knowledge runs counter to the scientific knowledge. And at still other times children have no clear common sense knowledge about the cultural/hrstorical knowledge we want to impart to our children via schooling. In all three situations we wiI1 probably need different teaching strategies to pass on the cultural/historical knowledge. Some of educational theory should be trying to tiork out how one bridges and connects between the common sense and cultural/historical knowledge so that they inform and support each other. And some of cognitive development theory (which does not now do SO) should be dealing with transitions within and transactions between the structures underlying children’s common sense and culturdl/historical knowledge (Vygotsky, 1962, 1978). In addition to the societal-educational view I just presented, there are other approaches that can be used to describe aspects of the problem I just presented. Most prominent among them that have informed my developmental work are structuralism a.nd information processing. There is a tendency today to view structuralism and information processing as alternative approaches. It is surely the case that some of the stiffest opposition to Piaget’s structuralism has come from those who ascribe to and work within one or another of the various information processing models. And this is probably as it should be since the struggle of ideas is often seen as an adversary relation in which each side tries to find weaknesses in the other’s position and to fight off opposition attempts to do the same. My work in both approaches (Strauss, 198 1; Strauss and Levin, in press) leaves me with the distinct feeling that they do not have to be alternatives and that they can be complementary, depending on the purposes behind one’s work. I believe that the two approaches appear to be different because of the level of derail thev are working at. The structuralist position, as outlined by Piaget, takes a rather broad stroke view of cognitive development whereas information processing approaches generally involve fine-grained analyses. Also, their purposes have often been different. Structuralism has reserved much of its energy for describing and elaborsting on the nature of mental organizations (structures) and occasionally about how they transform themselves. Information processing approaches often deal with operating rules about how information gets represented and transformed. But notice that the rules operate within an organization and this organization can be compatible with Piaget’s notion of structure. For example, Guy Cellerier from Geneva, Seymour Papert from MIT, and Juan Pascual-Leone from York University have all used different information processing approaches based on Piaget’s model of structuralism. There are others, of course, who ascribe to different information processing models and who would surely claim that they are not structuralists. And, indeed, this is probably the case. However,
the point I wanted to make here is that the information processing and structuralist models do not inevitably and inexorably lead to different places. In the end we might find out that they do, but as things stand now they can be seen, in some cases, as complementary. And now for my briefest statement of all. I believe that a deep problem btxtting the study of cognition is that our contemporary world of literature, theatre, pitiosophy, etc., does not have a clear image or metaphor of man as there once was in, say, the 18th and 19th centuries. This lack has a negative influence on the sort of theoretical and conceptual work being done in cognition and in education, as well. References Bamberger,J. (In press) Revisiting children’s draw&s of simple rhythms. In S. Strauss (ed.), U-shaped behavbralgrowth. New York, Academic Press. Carnap, R. (1966) philosophicrrlfoundations of physics. New York, Basic Books. Cohen, M. R. and Nagel, E. (1934) An introduction tc; logic a,+xdscientific method. New York, Harcourt, &ace, and Work!. Fel&nan, D. H. (1980) Bqond universalsin cognitive development. Norwood, NJ, Ablex. Hempel, C. G. (1952) Internationalencyclopedia of unified science, Vol. 2, No. 7: Fundamentals of concept formation in empirical science. Chicago, University of Chicago Press. Shef:ran, L. (1973) Vertical decalage in weight conservation between sensorhnotor and conceptual fevels. Unpublished honors thesis, University of Edinburgh, 1973. Reported in Bower, T. G. R. Concepts d development. In ploceedings of the 21s~ Inr.ernationalGmgress of Psychology. &is. PressesUniversitairesde France. Stavy, R and Berkovitz, B. (198U) Cognitive conMct as .I basis for teaching quantitative aspects of the concept of temperature. Sci. Educat...64,679-432. Strauss, S. (1981) Us&aped behaviora growth and eArcat5oa,Final Report for the Israeli Ministry of EdWltiOll. S?mas,
S. (In psess) Ancestral and descendant behaviors: The case of U-shaped behavioral growth. In T. G. &ever(ed.), Drops in leoraring.Hiusdale, NJ, Erlbaum. S*muss, S. and L&n, 1. (In press) Comments on Siegler. Monographsof fhe Society for Research in Gild Development. Strauss, SWaad &WY. R (In’press) U-shaped behavior%1 growth: Implications for theories of develop ment. In W. W. Hartup (ed.), Review of child development research Vol. 6, Chicago, University OfchicagQ Press. Strauss, S., Stavy R, and Orpaz, N. (1977) The child’s development of the concept of temperature. Unpu%lishedmanuscript. Tel-Aviv University. VYgpfelEy,l.. S. (1%2) Srhoorgltt and hmguage. Cambridge(Mass.), MIT Press. VYgotsky. L S. (1978) Mind tr bociety: The development of higher psychologicalprocesses. M. Cole, V. John-Steiner, S. Scribner, and E. Souberman @is.), Cambridge (Mass.), HarvardUniversity Press.
Cognition, 10 (1981) 301-306 @ Elrevier Sequoia S.A.., Lausanne - Printed in The Netherlands
301
The emergence of phonetic structure MICHAEL STUDDERT-KENNEDY* Queens College and Graduate Center City University of New York and Haskins Laboratories.
The earliest claim for the special status of speech as an acoustic signal sprang from the difficulty of devising an effecticjc alternative code to use in reading machines for the blind. Many years of sporadic, occasionally concentrated, effort have still yielded no acoustic system by which blind (or sighted) users can follow a text much more quickly than the 35 words a minute of skilled Morse code operators. Given the very high rates at which we handle an optical transform of language, in reading and writing, this failure with acoustic codes is particularly striking, Evidently, the advantage of speech lies not in the modality itself, but in the particular way it exploits tne modality. What acoustic properties set speech !.n this privileged relation to language? The concept of ‘encodedness’ was an early attempt to answer this question (Liberman, Cooper, Shankweiler and Studdert-Kennedy, 1.967). Liberman and his colleagues embraced the paradox that, although speech carries a linguistic message, units corresponding to thos;: ~>fthe message are not to be found in the signal. They proposed that speech should be viewed not as a cipher on linguistic structure, offering the listener a si.gnal isomorphic, unit for unit, with the message, but as a code. The code collapsed the phonemic segments (consonants and vowels) into acoustic syllables, so that cues to the component segments were subtly interleaved. The function of the code was to finesse the limited temporal resolving power of the ear. We typically speak and comfortably understand speech at a rate of lo- 15 phonemes/second, close to the rate at which discrete elements merge into a buzz. By packaging consonants and vowels into syllabic units, the argument went, we reduce this rate by a factor of two or three and so bring the signal within the resolving range of the ear. This complex code called for specialized decoding mechanisms. More than a decade of research was devoted to establishing the existence of a specialized phonetic decoding device in the left cerebral hemisphere and to *I thank Alvin Liberman, Ignatius Mattingiy and Bruno Repp far much fruitful discussion and ad&e. Preparation of this paper was supported in part by NICHD Grant HD 01994 to Maskins Laboratories. Reprint requests should be sent to Michael Studdert-Kennedy, Haskins Laboratories, 270 Crown Street, New Haven, Ct. 06510, U.S.A.
302
Mehad Studdert-Kennedy
isolating the perceptual stages by which the supposed device analyzed the syllable into its phonetic components. This information-processing approach to speech perception exploited a variety of experimental paradigms that had seemed valuable in visual research (see Darwin [ 19761 and StuddertKennedy [ 1976, 19801 for reviews), but led eventually to a dead end, as it gradually became apparent that the undertaking was mired in tautology. A prime ex; ,nple was the proposal to ‘explain’ sensitivity to features, whether phonetic Jr acoustic, as due to featuredetectjng devices, and to look for evidence oi such mechanisms in infants. Current research has drawn back’ and is now moving along two different, though not necessarily divergent, paths. The first bypasses the problems of segmental phonetic perception and focuses on what some believe to be the more realistic problem of describing the contributions of prosody, syntax and pragmatics to understanding speech. The second path, with which I am concerned, reverse; the procedure of the earlier encoding approach. fnstead of assuming that linguistic units should somehow be represented as segments in the signal and then attempting to explain the paradox of ,;neir absence by tailoring B perceptual mechanism for their extraction, the new approach simply asks: What information does the speech signal, in fact, convey? If we could answer this question, we might be in a position not to assume and impose linguistic structure, but. to describe how it emerges. Consider the lexicon of an average middle-class American child of six years. The chiId has a lexicon of some 12- 15,000 words, most of them Beamed over the previous four years at a rate of 7 or 8 a day. What makes this feat possible? Of course, the child must want to talk, and the meanings of the words he learns must match his experience: cut and funny, say, are more likely to be remembered than trepan and surd. But logically prior to the meaning of a word is its physical manifestation as a unit of neuromuscular action in the speaker and as an auditory event in the listener. Since the listening child readily becomes a speaker, even of words that he does not understand, the sound of a word must, at the very least, carry information on how to speak it. More exactly, the sound reflects a pattern of changes in laryngeal posture and in the supralaryngeal cavities of the vocal tract. The minimaI endowment of the child is therefore a capacity to reproduce a functionahy equivalent motor pattern with his own apparatus. What properties of the speech signal guide the child’s reproduction? We do not know the answer to this question. We do not even know the appropriate dimensions of description. But several lines of evid-ence suggest that the properties may be more dynamic and more abstract than cus’romary descriptions of spectral sections and spectral change. For example, some half
The emergence of phonetic structure
303
dozen studies have dtimonstrated ‘trading relations’ among acoustically incommensurate portions of the signal (e.g., Fitch, Halwes, Erickson and Liberman, 1.980; Liberman and Pisoni, 1977; Repp, Liberman, Eccardt and Pesetsky, 1978. Perhaps the most familiar example is the relation between extent of first formant transition and delay in voicing at the onset of a stop consonant-vowel syllable: reciprocal variations in spectral structure and onset delay produce equivalent phonetic percepts (Summerfield and Haggard, 1977). Presumably, the grounds of this and other such equivalences lie in the articulatory dynamics of natural speech, of which we do not yet have an adequate account. (For review of studies of this type, see Repp [in preparation] ) A second line of evidence comes from studies of sine-wave speech synthesis. Remez, Rubin, Pisoni and Carrel1 (1981) have shown that much, if not all, of the information for the perception of a novel utterance is preserved if the acoustic pattern, stripped of variations in overall amplitude and in the relative energy of formants, is reduced to a pattern of modulated sine waves following the approximate center frequencies of the three lowest formants. Here, it seems, nothing of the origin,al signal is preserved other than changes, and derivatives of changes, in the frequency positions of the main peaks of the vocal tract transfer function (cf. Kuhn, 1975). Finally, several recent audio-visual studies have shown that phonetic judgments of a spoken syllable can be modified, if the listener simultaneously watches a video presentation of a face mouthing a different syllable: for example a face uttering [gal on video, ,while a loudspeaker presents [ baj , is usually judged to be saying [da] (McGurk and MacDonald, 1976; Summerfield, 1979). The phonetic percept, in such a case, evidently derives from some combination of abstract, dynamic properties that characterize both auditory and visual patterns. Moreover, infants are sensitive to dynamic correspondences between speech heard and speech seen. Three-month old infants lcok longer at the face of a woman reading nursery rhymes, if auditory and visual displays are synchronized, than if the auditory pattern is delayed by 400 milliseconds (Dodd, 1979). This finding evidently reflects more than a general preference for audio-visual synchrony, since six-month old infants also look longer at the video display of a face repeating a disyllable that they hear (e.g. ilulul ) than at the synchronized display of a face repeating a different disyllable (e.g. [mama] ). (MacKain, StuCdert-Kennedy, Spieker and Stern, Reference note 1). The point here is not the cross-modal transfer of a pattern, which can be demonstrated readily in lower animals. Rather, it is the inference from this cross-modal transfer, and from the other evidence cited, that the speech
304
Michael Studdert-Kennedy
signal conveys information about articulation by means of an abstract (nnd therefore modality-free) dynamic pattern. The infant studies hint further that the infant learns to speak by discovering its capacity to transpose that pattern into an organizing scheme for control of its own vocal apparatus. Here we should note that, while the capacity to imitate general motor behavior may be quite common across animal species, a capacity for vocal imitation is rare. We should also distinguish social facilitation and general observational lea.rning from the detailed processes of imitation, evidenced by the cultural phenomenon of dialects among whales, seals, certain songbirds and humans. Firtally, we should note that speech (like musical performance and, perhaps, dance) has the peculiarity of being organized, at one level of execetion, in terms of a relatively small number of recurrent and, within limits, interchangeable gestures. Salient among these gestures are those that correspond to the processes of closing and opening the vocal tract, that is to the onsets (or offsets) and to the nuclei of syllables. We do not have to suppose that the child must analyze adult speech into features, syllables, segments or even words, before he can set about imitating what he has heard. To suppose this would be to posit for speech a mode of development that precisely reverses the normal (phylogenetic and ontogeaetic) process of differentiation. And, in fact, the earliest utterances used for symbolic ,or communicative ends, seem to be prosodic patterns which retain their unity across a wide variety of segmental realizations (Menn, 1976). Moreover, the early words also seem to be indivisible: for example, the child commonly pronounces certain sounds correctly in some words, but not in others (Menvuk and Menn, 1979). This implies that the child’s first pass at the adult model of a word is an unsegmented sweep, a rough, analog copy of the unsegmented syllable, And there is no reason to believe that the child’s percept is very much more differentiated than his production. Differentiation begins perhaps, when, with the growth of vocabulary, recurrent patterns emerge in the child’s motor repertoire. Words intersect, and similar control patte.rns coalesce into more or less invariant segments. The segmental organization is then revealed to the listener by the Lsild’s distortions. Menn (1978, 1980) describes these distortions as the result of systematic constraints on the child’s output: the execution of one segment of a word is distorted as a function of the properties of another. She classifies these constraints in terms of consonant harmony (e.g. [gAkJ for duck), conx:want sequence. (e.g., [nos] for snow;, relative position (e.g. [ daegel for ‘g@-~) and absolute position (e.g., [II] for fish). Here we touch on deep issues concerning the origin and nature of phonnlogical rules. Bllt the descriptive insights of Menn and others working in child phonology are important to the present argument, because they seem to justify a view of the phonetic segment as emerging from recurrent motor
The emergence of phonetic structure
305
patterns in the execution of syllables rather than as imposed by a specialized perceptual device. As motor differentiation proceeds, these recurrent patterns form classes, defined by their shared motor components--shared, in part, because the vocal tract has relatively i’ewindependently movable parts. These components are, of course, the motor origins of phonetic features (cf. Studdert-Kennedy and Lane, 1980). Some such formulation is necessary to resolve the paradox of a quasi-continuous signal carrying a segmented linguistic message. The signal carries no message: it carries information concerning its source. The message lies in the peculiar relation between the source and the listener, as a human and as a speaker of a particular language. Readers familiar with the work of Turvey and Shaw (e.g., 1979) will recognize that the present sketch of a new approach to speech perception owes much to their ecological perspective (as also to Fowler, Rubin, Remez and Turvey, 1980). What may not be generally realized is that this perspeclive is highly compatible with much recent work in natural phonology (e.g., Ylitampe, 19731, child phonology (e;g., Menn, 1980) and phonetic theory I:e.g., Lindblam, 1988; MacNeilage and Ladefoged, 1976; Ohala (in press). for exarnple, Lindblom and his colleagues have, for several years, been developing principles by which the feature structure of the sound systems of different languages might be derived from perceptual and articulatory constraints, More generally, Lindblom (1980) has stressed that explanatory theory must refer (. . . to principles that are independent of the domain of the observations themselves’ (p. 18) and has urged that phonetic theory ‘. . . move [its] search for basic explanatory principles into the physics and physiology of the brain, nervous system and speech organs.. .’ (p. 18). In short, if language is a window on the mind, speech is the thin end of an experimental wedge that will pry the window open. The next ten years may finally see the first steps toward a genuine biology of language.
Referexes Darwin, C. J. (1976) The ycxception of spexh. In Carterette, E. and Friedman, M. (eds.), Handbook of Perception, Vol. VI,‘. New York, Academic Press, 176-226. Dodd, B. (1979) Lip reading in infants: Attention to speech presented in- and out-of-synchrony. Cog. Psych&, II, 478-484. t’rickson, D. M. and Liberman, A. M. (1980) Perceptual equivalence of two Fitch, H. L., Halwes, T., L, acoustic cues for stop-cnnsonant manner. Percep. Psychophys., 27. 343-350. Fowler, C. A., Rubin, P., Remez, R. E. and Turvey, M. T. (1980) Implications for speech production of a general theoiy of action. In Butterworth, B. (ed.), Language Production. New York, Academic Press, 373-42. Kuhn, C. M. (1973) On the front cavity resonance and its possible role in speech perception. J. Acoust. Sot. Am., 58, 428-433.
306
Mkhuel Studdert-Kennedy
Liberman, A. M., Cooper, F. S., Shankweiler, D. P. and Studdert-Kennedy, M, (1967) Perception of the speer’ code. Psychol. Rev., 74,43 l-461. Liberman, A. M. and Pisoni, D. B. (1977) Evidence for a special speech-perceiving mechanism in the human. In Bullock, T. H. (ed.), The Recognition of Complex Acoustic Signals. Berlin, Dahlem Konferenzen, 59-76. Lindblom, B. (1980) The goal of phonetics, its unification and application. Phonetica, 37, 7-26. McGurk, H. and McDonald, J. (1976) Hearing lips and seeing voices.Nuture, 264, 746-748. MacNeiiage, P. and Ladefoged, P. (1976) The production of speech and language. In Carterette, E. C. and Friedman, M. (eds.), Handbook of Perception, Vol. VII. New York, Academic Press, 75-120. Menn, L. (1976) P&em. Control and Contrastin Reginning Speech: A t&e Study in the Develop menr of Word Fom mrd Word Function. Bloomington, Indiana University Linguistics Club. Menn, L. (1978) Phonological units in beginning speech. In Bei:, A. nd Hooper, J. B. (eds.), Syllables und Segments. Amsterdam, North-Holland. Menn, L. (1980) Phonological theory and child phonology. In Yeni-Komshian, G., Kavanagh, J. F. and Ferguson, C. A. (eds.), Child Phonology: Perceptionand Reduction, Vol. I, 23-41. Menyuk, P. and Menn, L. (1979) Early strategies for the perception and production of words and sounds. In Fletcher, P. and Garniar, M. (eds.), Lcnguuge Acquisition. New York, Cambridge University Press, 49 - 70. Ghala, J. (In press) The origin of sound patterns in vocal tract constraints. In MacNeiIage, P. F. (ed.), Speech Reduction. New York, Springer-Verlag. Remez, K. E., Rubin, P. E., Pisoni, D. B. and Carrell, T. D. (1981) Speech perception without traditional speech cues. Scienre 212. 947-950. Repp, B. H. (In preparation) Phonetic trading relations and context effe$ *s: New experimental evidence for a speech mode of perception. Repp, B. H., Liberman, A. M., Eccardt, T. and Pesetsky, D. (1978) Perceptual integration of acoustic cues for stop, : fricative, and affricate manner. J. Exper. Psychol.: Hum. Percep. Perf, 4, 621-637. Stampe, D. (In press) A Dissertationon NaturalPhonology. NeroYo:k, Garland. Studdert-Kennedy, M. (1976) Speech perception. In Lass, N. J. (ed.), ContemporaryIssues in Experi. mentalPhonetics. New York, Academic Press, 243-293. Studdert-Kerr&y, M. (1980) Speech perception. Lung. Sp., 23, 45-66. Studdert-Kennedy, M. and Lane, H. (1980) The structuring of language: Clues from the differences between signed and spoken language. In Sellugi, U. and Studdert-Kennedy, M. (eds.), Signed Language and Spoken Language: .RiologicalConstminrson LinguisticForm. Deerfield Beach, Fl., VerIag Chemie, 29-410. Summertield, Q. (i979) Use of visual information for phonetic perception. Phoneticu,36, 314-331. Summer~eld, Q. and Haggard, M. (1977) On the dissociation of spectral and temporal cues to the voicing distinct!Jn in inithl stop consonants.J. Acoust. Sue. Am., 42, 436-448. Turvey, M. T. and Shaw, R. E. (‘1979) The primacy of perceiving: An ecological reformulation of perception for understanding memory. In Nilsson, L.-G. (ed.), Perspectives on Memory Research: Essays in Honor of Uppsah’s 500th Anniversary. Hill&dale, NJ, Erlbaum.
Reference Note
MacKain, K., Studdert-Kennedy, M., Spieker, S. and Stern, D. Cross-modal r3ordination in infants’ perception of speech. Paper to be read at the Second International Conference on Child Psychology, Vancouver, B.C., August 1981.
Cognition, 10 (1981) 307-312 @ Ekvier Sequoia S.A., Lausanne - Printed in The Netherlands
307
The process of language comprehension; an approach to examining issuesin\ cognition and language DAVID SWINNEY* Tufts University
The ftmdamental concer.ns of the field of cognitive ps;.chology-understanding the nature of the mental representation of information and the processes which operate on those representations--have received their most extensive examination in the past decade in the field of psycholinguistics. The reasons for this are obvious: language is a relatively accessible domain for study, and the systematic classification and observation brought by linguistics and philoso?hy to language provided a major initial stepping stone for the investigation of the psychological functions underlying human language production and comprehension. And, \vhile there has been much growth in a number of other fields in the area of cognition, the now-expanded domain of psycholinguistics still represents one of the most l_romising and profitable areas foi examining cognitive function. It appears likely that major breakthroughs in understanding the extent of (and constraints on) our ability to perceive, process, store, recall and comprehend information will come from gaining a detailed empirical understanding of the nature of language processing. It is particularly important in this regard to underscore the poi+ that language, like other cognitive functions, is a (dynamic) pwtxss. It is, m fact, precisely because it is a process that gaining an understanding of its nature has proved so intractable over the years; such an enterprise requires that we rely on more than, just an examination of the end-state characteristics of the process (such as memory representations) or static models of its putative underlying structure. Rather, if we are to achieve any sublrtantive understanding of language performance it is necessary that we examine the microstructure of the entire process as it occurs in real time. It is only through the careful examination of the temporal course of mental operations involved ir. the various levels of analysis underlying speech that we can hope to discover its nature. There are a number of critical questions underlying the examin.ation of language as a process which must be answered in the coming years. The first, and in one sense most fundamental, concerns the basic nature of the integra*Reprint requests should be sent to: David Swinney, Psychology Department, Tufts University, Medford, Mass.02155, U.S.A.
tion of information that occurs duriklg the perceptual processing [or production) of language. This issue has evolved in the past few years into a hotly contested one, one that contrasts hypotheses that, on the one hand, view language processing as a maximally interactive system (in which any type of contextual information can affect the nature of processing of any other piece of inf jrmation) and, on the other hancl, view the system as a highly modular orle comprised of autonomous subroutines (so that conrextual information does not affect processing internal to any particular subroutine). Resolution of the issue of whether the s:rstem in general is a modular or an interactive one is perhaps the most impo.stant key to establishing a viable psychological model of perceptual processulg. Certainly, it is clear that a system wh’sh supports the nearly infinite variability in processing that is inherent in any maximally interactive model requires a significantly more powerful underlying mechanism th*an one which consists of relatively autonomous, context free, internally consistant routines. Failure to provide clear definition of tile nature of information integration almost guarantees that we will posit either far tDo powerful or far too weak an underlying mechanism as the basis for language behavior, a failul 9 which may well be a critical one in our atten qts to understand the procebs. However, in addlcion to being important in and of itself, part of the importance of focusing on the nature of .nformation integration is that it highlights several other key problems that must be faced in discovering the nature of language processing. One set of these issues concerns the nature of thr. informationai types themselves. We do not have an account of the didtirlct types of informatior that are actually functional in speech processing. While systematic observation in linguistics has allowed for the char;cterization of a number of sources of information 1 language-phonetic, morphemic, syntactic, etc. -such descriptions are based un the rules and constraints of linguistic descripti,. n and have yet t,o be shown to have anything but an indirect relationship to the types of information which are functional in the ongoing process SC language comprehension. The development of the psychologically relevant characterization of the functional sources of imarmation ilj language processing is a critical goal for psycholinguistics, and it is important to note that it is a problem that can only be resolved by empirical study. A fidrther, related, issue stems from the fact that the characterizations we give tc both whatever relel ant informational types exist and to the procedures by which these functic.,n are intimately tied to assumptions we make about the techniques by which we examine these prucesses. There is simpiy no passive window which’ allows examination of mental processes without affecting those processes to some extent. Thus, it is absolutely essential at
Lie process of language camp-chension
309
this time that we begin to consider the modeling of our empirical tasks as being a critically important part of our attempt to model mental representations. It is encouraging to note that, from time to time, various experimental techniques have undergone critical examination (witness the study of phoneme monitoring over the past nine year s). However, far too little is known of even the most basic characteristics of those tasks which have attracted such attention, not to mention those that have been ignored in this respect. There are at least two corollary issues that are of importance here. The first relates to the critical need for discovering, testing, and using new experiAmentaltechniques. Given that the problems we face are to be best resolved by empirical examination of the real-time characteristics of language processing, we are in need of experimental techniques which are sufticiently flexible to examine such processing on-line and, simultaneously, which are as non-intrusive into the process under study as possible. Further, we are in great need of tasks which are differentially sensitive to various levels of analysis in language processing. It is important that examination of mental processes be seen to be a multi-leveled enterprise. We need to discover the nature of c :h of the several putative levels of processing as they occur (either serially or in parallel) during the on-going perceptual analysis of language; the relative speed, interactivity, and scope of each of these processes must be detailed. Thus, the relative temporal course of processing, the accumulation of detail, and the r&e of integration of each process into the developing interpretation needs to be documented for (for example) Fhonetic, lexical and syntactic information during comprehension. To do this, it will be necessary to develop batteries of ‘on-line’ tasks, each ot which has known properties that reflect different aspects of the process under study. The second corAary issue concerns the need for determining the relationship that holds between conscious and unconscious processing and automaticity of processing. There is not sufficient space here to even begin to detail the degree to-which these interrelated issues are critical to the enterprise of understanding language processing. However, because language is a highly OVIrlearned, automatized system, until we understand the consequence; of that fact and the way in which it affects our ability to interrupt, examine, introspect upon, have intuitions about, and manipulate language, we will have little idea of the tr.:a information contained in the data tha.t we are gathering. Similarly, until M2 understand the cognitive operations un.derlying the ability to bring the results of unconscious processing (the level of most of language processing) to consciousness we will not have adequate knowledge of how our exper.l.lental tasks may be changing the basic 01~~ation of the language processing system. Further, gainir‘g an understanding of consciousness and automaticity will provide a basis for examining the nature
of the relationship between putative language processes and more generalized, domain-universal, cognitive functions. The distinction between such processes is clea.rly a critical one in our development of cognitive processirg theories. Specific approaches
The immediate goal of much of our work over the past six years has been to discover the nature of the processes that underlie language comprehension through the careful examination of the information available throughout the course of sentence/discourse understanding. We have taken the major testing ground for examination of this general theoretical issue to be the domain of lexical processing. This choice has been made for a number of reasons. First, wordls are likely candidates for being truly functional sources of information during language processing; word recognition is acknowledged to play a role in nearly every psychological account of language, just as words are taken to be important structures in most linguistic accounts. In addition, lexical representation and processing are commonly considered to be (the) major points of intersection of acoustic-phonetic, syntactic, semantic and discourse i.nfoimation in language. Thus it is a logical realm in which to examine the general nature of information interaction, the (relative) temporal characteristics of language processing, and theoretical issues of automaticity, uncons.cious inference, and the relationship of linguistic and psychological modeling. An important phase of our work has involved the examination of the maximally interactive and autonomous module hypotheses of information integration as they apply to the access and processing of lexical material. We have studied the effects of a large range of ‘higher order’ contextual information sources (from local lexical semantic and structural relationships to more global sentential and discourse cjffects) upon ths access and integration of lexical material in ongoing sentepce/discourse comprehension. In pursuit of this work, we have developed a relatively sensitive, nonintrusive, flexible, measure of the unconscious access of lexical information -the cross modal lexical priming task. The work :!sing this task has been illuminating in many regards. We have, throughout a number of studies, found strong support for the hypothesis that (at least) lexical access is an autonomous, form-driven process, one that operates without regard to biasing input from higher order contextual information (e.g., Swinney, 1979; Swinney, Onifer, Prather, and Hirshkowitz, 1979; Onifer and Swinney, 198 1; Swinnt~y 1981a; 198lb). T!~ough the examination of the temporal course of activatiion of lexical material we have also learned much about the nature of the po@access effects of frequency-of-meaning, type of contextual information, and speed of processing upon the choice of a relevant interpretation for
The process of language comprehension
3 11
a word from among all of the information that is initially acck:ssed. Some information about the roles that automaticity and consciousness play in the context-independent process of access and in the contextually incorporative post-access decision proce% have also begun to be determined through the use of cross modal priming and other on-li:.e procedures (Swinney, 198 lb). Some of the most e&citing evidence, however, has come from our examination of these real-time processing issues in populations which have specific language disorders, disorders which provide theoretically dissociative test cases for study of the issues raised above, and in immature populations, populations in which the highly automatized routines which exist in adult language processors have yet to develop. The study of the on-line processing of words, structures, and suprasegmental informati:>n in young children and in populations with aphasias, schizophrenia, and dy sslexiasare providing critical diagnostic evidence for the validity of the modular (autonomy) theory of languafle processing as well as providing critical evidence about the primacy of certain types of these operations and about the nature of the automatization of such routines in language (see, e.g., Swinney, Zurif and Cutler, 1980; Swinney, 1981b). P’Drk involving such a multidisciplinary approach seems to us to be absolutely necessary for the development of a firm understanding of mental operations, if only becar:se it is often just in the absence of overlearned routines and fully functional systems that some of the microstructure of mental events become apparent. An additional, and we think important, focus for much of our recent effort has been in the relatively neglecttid area of non-literal language processing. While it is understandable that much of the early work in language focused solely on the processing of literal material, it is equally clear that an extremely large amount of the language that we deal with is non-literal: idoms, similies, metaphors, and the like. .C)ur major efforts in the past two years have been on attempting to understand the on-line processing characteristics of nonliteral as well as literal processing-to determine whether the processor attempts such. interpretations (particularly where multiple interpretations are possible) by use of a cannonical order hypothesis, by reliance on surface cues, or by some other means. The evidence thus far (see, e.g., Swinney and Cutler, 1979) both supports the modular theory of perceptual information integration and adds much evidence to our general understanding of the role of learning in language perception. Finally, much of our recent efforts have been aimed toward the examination of the underlying properties of semantic and syntactic priming, both in discourse contexts and in isolation. The results have suggested ihat much of what has been assumed to be an ‘automatic’ spread of facilitation in prim@ may in fact be the result of a relatively elaborate decision process. Recent
312
David Switmey
work with Penny Prather has extended these findings into both the area of speech perception and the domain of visual processing, in an attempt to discover the degree to which much of the evidence supporting a processor comprised of independent subsystems applies to other domains of cognitive processing. In all, the approach that we have taken in examining many of the questions raised above -the nature of information interaction, the types of information that are functionally availdble in processing, the temporal course of conscious and unconscious processes, and (overall) the nature of the reof language-has obviously presentations involved in the perceptual anal) merely scratched the surface of the complexity of both theory and fact that must be developed to provide a sufficient characterizat.ion of the cognitive system. However, we have found it particularly instructive to examine the form of the evidence that has emerged from an extensive examination of one small portion of the language system -lexical processing. Particularly because this evidence is in such agreement with that obtained from our broader, multidisciplinary examinations of language processing, we take it . be likely that it is representative of the nature of cognitive processing in gcneral. Independent of the validity of this interpretation, it is important to reemphasize that the issues are largely empirical ones, and thus we can look forward to their resolution through the development of well reasoned experimental evidence in the coming years. References Keeie, B., and Swinney, D. (1979) On the relationship of hemispheric specialization and develop mental dyslexia. cortex, 15.471-481. Chifer, W., and Swinney, D. (In press) Assessing lexical ambiguities during sentence comprehension: Effects of frequencyof-meaning and contextual bias. Mem. chg. Swinney, D. (1979) lexical access during sentence comprehension: (Re)consideration of context effects. i. verb. Learn. verb. Behlrv., 18, 645-660. Swinney, D. (In press a) The structure and timecourse of information interaction during speech comprehension: Lexical segmentation, access, and interpretation. In Mehler, J., Garrett, hf., and Walker, E. T. C. (eds.), h>ceedings of the Royaumont cbnference on Cbgnitive fiychology. Swinnw, D. (In press b) Jxxical processing during sentence comprehension: Effects of higher order constraints and implications for representation. To appear in T. Meyers, J. Laver,and.J. Ander?on (eds.), TIIe C@nitive Representation of Speech (Advances i. Psychology Series). Amsterdam, North-Holland. Swinney, D., and Cutler, A. (1979) The access and processing of idiomatic expressions. J. verb. Learn. verb. Behav., 18.523-534. Swimmy, D.,Gnifer, W., Rather, P., and Hirshlcowitz, M. (1979) Semantic facilitation across sensory modalities in the proces&g of individual words and sentences, Mem. Chg., 7, 154-165. Swinney, D., and Rather, P. (1980) Phoneme identification: the role of within-syllable context in monitoring for syllableinitial consonants. Pereep. &vchophys., 27,104-l 10. Swirmey, D., Zurif, E., ard Cutler, A. (1980) Effe:ts of sen&&al stress and word class upon comprehension in Broca’s aphasics. Brain Lung., 10, 132-144.
Cognition, 10 (1081) 313-321 @ Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
Cognition: The view from ecological
313
r
alism
M. T. TURVEY* University of Connecticut and Hask, :” L id”, 3ra tories CLAUDIA
CARELLO
University of Connecticut
The term ‘cognition’ is taken, very generally, to refer to the cocxdination of any organism (as an epistemic agent) and its environment (as the support for its acts). The task of cognitive theory is to explain this epistemic, intentional coordination of organism and environment. Orthodoxy subscribes to the Lockean view that the coordination is achieved through, and explained by, a special class of things. Locke called these things that interface organism and environment ‘ideas. ; contemporary cognitive theorists lean toward ‘representaticns’, ‘programs’, ‘reference signals’, ‘schemata’, etc. Each of these coordinating things is an entity presumed to be endowed with properties that are (sometimes loosely, often d-ictlyj isomorphic with those properties of the state of affairs for which the coordinator is said to be causally responsible. Moreover, it is also presumed that a coordinating thing is of the same logical t:?rpe as the organism, snvironment state of affairs that it putatively explains- each ir an instance of intelligence (or knowing, or rationality, or goal-directedness). For the orthodox theorist, an appropriate candidate to play the role of coordiriating thing in a given situation is arrived at by inference: Those properties and that organization that are sufficient to describe .the observed phenomenon must be identified. The assumption is that actual coordinating things are determined similarly, that is, that these instances of i:ntelligence are arrived at intelligently. This view is troublesome on two counts, neither of which seems to bother establishment theorists. First, orthodox cognitive theory is not distressed by the large loans of intelligence that its program demands. Given that coordinating things are rational entities arrived at by rational means, how does orthodox theory account for the ultimate origin of all this rationality? Second, the orthodox view is similarly unconcerned with the skepticism engendered by the assumption that perception is a relation of an organism to an internal representation of its environment. Given that inference can fail, and that the cm*This paper was written while the first author was a Fellow zt the Center for Advanced Study in the Behavioral Sciences. Support from NSF Grant BNS 76 22943 is gratefully ‘acknowledged. Reprint requests should be sent to: M. Turvey, Department of Psychology, University of Connecticut, Storrs, Connecticut, 06268, U.S.A.
314 M. T.T~tyamiCCardlo
sequences of inferences unfettered by real constraints are vacuous, how does orthodox theory insure that an inference-determined representation represents an actual state of affairs, that is, that perception is not of fictions but of the real environmental ,things (types of substance, surface, place, object ,and event) with respect to which acts are conducted? The second concern-skepticismcan be alleviated in the orthodox perspective, but only by exacerbating the first---taking out further loans of intelligence. Investing an organism with detailed foreknowledge of the conditions of its ecosystem might build into orthodox theory sufficient constraint to guarantee that the organisn’s inferences are realistic. However, to presuppose the very thing that is to be explained is a move that no serious science of cognition can abide. It seems, therefore, that the orthadox approach to cognition is seriously infirmed 6;ld that the foundations of cognitive theory demand a radical rethinking of the kind initiated by Gibson (1950, 1966,1979). The heterodox, ecological approach explicitly recognizes ‘knowing’ as a natural phenomenon at a particular scale of magnitude, viz., the ecological scale of living things and their niches. It requires that ontology and epistemology-and the scientific theory and method that they shape-be tailored to the ecological scale ‘(Gibson, 1979 ; Michaels and Carello, 198 1). It rejects any strategy that inputes coordinating things to explain cognition and embraces, instead, a strategy that searches for natural laws at the ecological scale (fashioned by scale-independent principles) that coordinate organism and environment (Turvey, Shaw, Reed and Mace, 1981). The ecological strategy observes two rules of thumb: (1) resist taking out loans of intelligence; and (2) regard with skepticism, and be prepared to jettison, any assumption, concept, interpretation, fact, theory, strategy, etc., that undercuts or threatens to undercut the principle of ecological realism. This principle can be sketched roughly as follows. An activity of an organism is a nesting of behavioral adjustments to a nesting of environmental properties. (For exampr_e, a-bird searching for insects is oriented at a fine spatiotemporal scale to the.crevice at which it directs its beak, at a less fine spatiotempora! scale to the branch on which it stands, and at a much coarser spatiotemporal scale to the sky-earth light differential). To control its activity, the organism must perceive both the nested environmental properties and its own nesf4 behaviors. The principle of ecological realism is the assertion that the nested pro.perties of organism and environment are objective states of affairs of the organism-niche system (that is, their existence is independent of whether ~1 not the organism experiences them) and it is these states of affairs, and ody these, that are the ‘objects’ of perception (Gibson, 1979; Shaw, Turvey and Mace, in press; Turvey and Shaw, 1979). Any move that will inevitably replace the aforementioned ‘objects’ of perception with some others (pop-
Cognition: The view frotm ecological realism
315
ularly, neural states and/or mental states) undercuts the principle and, therefore, is not allowed. The commitment to the principle of ecological realism expressed in rule (2) she. Id not be undervalued. In our view, that commitment directs the ecological program and makes it cohere. A renunciation of realism would make a mockery of what scientists do qua scientists (Bunge, 1973; d’Espagnat, 1979) and a renunciation of the principle of ecological realism would make a mockery of what scientists do qua organisms (and what organisms do in general). Without the principle, the adaptive relation of organism and niche would be in the style of the cartoon character Mr. Magoo who never sees what behaviors the environment actually permits nor what behaviors are actually performed but who survives nevertheless through the benevolence of the animator. The adaptive relation of organism and niche over the short term (popularly referred to as perceptuo-motor coordination) requires the principle of ecological realism, as do the adaptive relations over the medium t\;lm (learning), and the long term (evolution) (see Johnston and Turvey, 1980). The principle of ecological realism, therefore, is taken to be the fundamental principle of a science of cognition. In our view, while types of inquiry other than scientific may not be obliged to preserve the principle at all costs, science is obliged. For a science of cognition, however, it is an obligation that is not met easily-cognitive theory feeds on various disciplines that, themselves, not only contain many(entrenched) conceptions and methodologies that deny the principle but also lack other conceptions and methodologies needed to sustain it. The principle’s ultimate significance is that,for any discipline,it picks out concepts, assumptions, interpretations, etc. that are unacceptable and it points to those concepts, etc. that are required, whether they be currently available or not. The ecological approach has identified a number of such conceptions, etc. Some of athe major ones are listed below,. togeth.er with a brief descriptiort (most have been advanced in some detail elsewhere). Collectively, these conceptions define a framework in which to investigate cognition.
1. Organisrn-environmcat relation Perhaps fFe fundamental conception of ecological realism concerns the logical dependence of organism and environment (Gibson, 1979; Turvey .and Shaw, 1979). Organism-environment synergy does not suggest merely that an organism implies the existence of some environment (or vice versa) but, more strongly, that each component of an organism-niche system logically conditions the very nature of the other component (Patten, 1979; Shaw and Turvey, 198 1). Such a claim demands that organism-niche systems be the
316
hf. I.7Tuweyand C tirello
irreducible units of analysis for understanding the phenomena of knowing (Michaels and Carello, 198 1; Shaw, et al., in press). 2, Description of the environment In order to be consistent with the principle of ecological realism, a useful description of the environment must reflect an environment’s mutually constraining relationship to the organism. Such an organism-referential descrip tion is provided in Gibson’s concept of affordance, a tiescription that captures the nature of a given niche as the environmental support for a particular animal’s activities (Gibson, 1979). The behavioral concerns motivated by the principle of ecological realism are currently being brought to bear on the concept of niche in zoology and ecology (Alley, 1981), where it has been conceded that traditional treatments of physical and biological needs alone are incomplete (Whittaker, Levin and Root, 1973). In highlighting the mutuality between an organism’s activities and the niche in which that organism evolved and with respect to which those activities are adaptive, the efficacy of borrowing organism-neutral phlysical taxonomies and simplistic biological taxonomies is denied (Runeson, 1977; Turvcy and Shaw, 1979; Turvey et al., 1981 j. Appropriate physical and biological taxonomies have yet to be determined but it is contended that the principle of ecological realism will be a necessary constraint on their ultimate formulations (Shaw and Cutting, 1980). 3*ll&ormation Neither classical information theory nor the currently popular quasilinguistic view of information is acceptable (Gibsor, 1966; Kugler, Kelso and Turvey, in press). The ecological approach asserts that the concept of information cannot be developed systematically apart from considerations of activity. In scientific discourse, ‘information is used in many contexts, but it is in the contexts of coordinating and controlling activity (more generally, dynamics) that its use is most pronounced and its nature most elusive. Ecological realism imposes severe demands on the concept: Information must be unique and specific to the facts about which it informs (Gibson, 1979 ; Mace 1977 ; Reed and .Iones, 1978; Turvey and Shaw, 19791, meaningful to the coordination and control requirements of the activity (what can be done, how it can be donr+ and when it can be dQne) (Gibson, 1979; Lee, 1980; Michaels and Carello, 1981; Fowler and Turvey, 1978), and continuously scaled to the dimensions of the system over which the activity is defmed (Kugler, et aT., in press).
Cbgnition: The view from ecological realism 3 17
4. Natural law The interpretation of natural law as a relation between classes of things (more formally, an intensional relation between extensions) denies specification in the sense of one property lawfully related to another, for example, an activityrelevant property of the environment lawfully related to a property of ambient light (Fodor and Pylyshyn, 1981; Turvey et al, 1981). As noted in (3), ecolsgical realism mandates specification. The ecological approach requires, therefore, that a natural law be a relation between properties (more formally, an extentional relation between intensions) rather than a relation between classes (Reed, 1979; Turvey, etal., 1981). 5. Units or scales of measure Given that extrinsic units of measure sustain animal-environment dualism and, therefore, undermine ecological realism, they are not tolerated. Very roughly, given that extrinsic units (meters, grams, etc.) are arbitrarily imposed on an environment, they require that the organism perform a complicated conversion in order to derive units appropriate for activity. Units that are intrinsic to an organism-niche system, on the other hand, share common bases in the organism and niche (.;;lhawand Cutting, 1980) such that certain parts and processes of the ;ysPem define the units in which other parts and processes are measured (see, for example, Lee, 1980; Sedgwick, 1973). The task of systematically determining intrinsic measures may be facilitated by the identification of constraints that make the coordination and cciltrol of movement possible. For example, a measurement system appropriate to the activ’~:ds of a living thing should consist of three coimplicative metrics-an intrinsic temporal metric, an i.ntrinsic spatial metric, and an intrinsic power metric-such that fixing the value of any two of them on any given occasion will naturally constrain the third (Kugler, personal communication). A paradigmatic instance is a tennis player who requires information appropriate to her dimensions for where to contact the ball, when contact is to be made, and how much power is required in ordry to arrive at the appropriate place at the, appropriate time.
6, Semantics Oirthodox semantics is rooted in (i) the analysis of formal mathematical languages, with its notions of meaning as a fixed property of an expression and an unrestricted universe as the ground for interpreting sentences, and (ii) a
318
M. T. Turcey and C. Gzrello
tradition that identifies the intentionalit:l of everyday language with a commitment to concepts or mental represer .ations, which can be representations of ‘inexistent’ things. From the ecorogical perspective, however, semantics begins with a commitment to the view that language is intentional because it reports on the relations of the user to the propertied things that make up the language user’s environment (e.g., BarsAse, in press; Barwise and Perry, 198 1) and not because it reports on concepts, ideas, etc., as is the orthodox claim. The interpretation of expressions, therefore, must be in terms of the propertied things and the relations that compose ihe speaker-environment system. And the context-free meaning of e\:,*yessions assumed in orthodox semantics must give way tc the ecological fact that the linguistic expressions that intimately link human and environment (such as ‘this’, ‘that’, ‘today’) assume different designations, that is, pick out different propertied things and relations, as a function of the user, event, and place (Barwise, in press; Barwise and Perry, 1981). 7. Physical theory
*
Until recently, physical theories have sought, by and large, to give causal accounts of transitions between states of the same order of complexity (mechanics) and have addressed, to only a limited degree, the question of how various orders of complexity might arise (equilibrium, reversible thermodynamics). The orthodox approach to cognition couples a tendency to assume that physical theory is complete with the supposition thitt the kinds of orders that characterize cognition are largely outside physics’ explanatory power. This coupling leads to the promotion of a special explanatory vocabulary, that of representation and computation. But, as implied above, coordinating things are prescriptions for states of affairs and come very close to embodying the very order that they are meant to explain; and a discrete, symbolic mode yields a vocabulary that is proprietary only for explaining an ideal system existing in an ideal universe in which space, tinlti, matter, and energy make no contribution (Pattee, 1974; Shaw and Rlclntyre, 1974). Commitment to the principle of ecologizaI realism cautions against as suming that physics is complete and, instead, advocates patience with regard to physics’ eventual contribution to cognition. It underscrores the need to ground cognitive theory in a physical system that is real rather than ideal. Given that the principle requires that epistemic, intentional states of an organism-niche system be a ‘posteriori facts, not a priori prescriptions (i.e., states that arise from the design of the organism-niche system rather than states that are imposed by a model of the organism-niche system) (Kugler, et al., 1980;
agnition:
The viewfrom ecological realism
3 19
Turvey, et al., 198 l), the charge to physical theory is an acccclnt of the physical principles of design that govern the evolution of different orders of complexity in a contirluous dynamical system. This charge is being met by physical theories that couple statistical mechanics and nonequ?brium irreversible thermodynamics (e.g., Iberall, 1972, 1977; Morowitz, 1978; Prigogine, 1978; Yates, 1980). This curious feature of an ecological approach-that the conceptions, etc. of the various branches of knowledge must be consistent with the principle of ecological realism- marks the ecological approach as an eccentric way of doing science. It asserts that, at the ecological scale, a certain principle must hold for the phenomena characteristic of that scale to be manifest, and it uses this principle- encouraged by a belief in the mutual compatibility of natural phenomena (Shaw and McIntyre, 1974; Shaw and Turvey, 198 1; Turvey and Shaw, 1979)-as a benchmark for evaluating statements about st;ltes of affairs at all scales of magnitude. Happily, this eccentricity is not limited to proponents of the ecological view. It is intuited in other circles, for example, that the basic features of the universe are understandable in terms of a few physical constants together with the constraint that the order ->fmagnitude of these constants be consistent with the fact of living things (Carter, 1974; Carr and Rees, 1979). The ecological approach expresses a similar intuition but one that we suspect is more far-reaching: A thoroughgoing explanation of perception in the service of activity, consistent with the principle of ecological realism, will impose powerful constraints not only on cognitive theory but on physical and biological theory as well.
References Alley, T. (1981) A re-examination of the concept of an ecological niche. Unpublished manuscript, University of Connecticut. Barwise, J. (in press) Scenes and other situations. J. Phil. Barwisc, J., and Perry, J. (1981) Semantic innocence and uncompromising situations. MidwestStudies in Philosophy, 6. Bunge, M. (1973) Philosophy ofphysics Dordrecht, Holland, D. Rsidel Publishing Co. Carr, B. J., and Rees, M. J. (1979) The anthropic principle and the structure of the physical world. Voture, 278,605-612. Carter, B. (1974) Large number coincidences and the anthropic principle in cosmology. In M. S. L0ng;d.r (ed.), Confrontation of cosmologicaltheories withohsercotionaldata.Boston, D. Reidel Publishing Company. cl’Espagnat,B. (1979) The quantum theory and reality. Sci. Amer., 241, 158-181. Fodor, 3. A., and Pylyshgn, 2. W. (1981) How direct is visual perception? Some reflections on Gibson’s ‘Ecological Approach’. Cog,, 9, 139 -196. Fowler, C. A., and Turvey, M. T., (1978) Skill acquisition: An event approach with special reference to searching for the optimum of a function of several variables. In -3. Stelmach ted.), Inforn:ntion piocessing Pr motor control and learning. New York, Aca:jemic Press.
Gibson, J. J. (1950) 77rep~rcePtion of the visual world, Boston, MA, Houghton-Mifflin. Gibson, J. J., (1966) The senses considered QSpemeptual systems, Boston, MA, Houghton-Mifflii. Gibson, 3. 3. (1979) The ecological appmch to visud perception. Boston, MA, Houghton-Mifflin. Iberall, A. S. (1972) Towni a general science oftible systems. New York, McGraw-Hill. Iberall, A. S. (I 977) A fieki and circuit thermodynamics for integrative physiology: 1. Introduction to general ncrtions. Am. J. Physh$Zeg. Int. Camp. Hysio.., 2, R171 -Rl8Q. Johnston. 1.. and Turvey, M. 1. (1980) A sketclr of an ecological methatheory for theories of learning. In G. H. Bower(ed.), f7Ire pathology of l%rningand motivation. (Vol. 14). New York, Academic Press. Kugler, P. N., Kelso, J. A. S., and Turvey, M. T. (1980) On the concept of coordinative structures as dissipative structures. I. Theoretical lines of convergence. In G. E. Stelmach and J. Requin (eds.), 7 rtorhds in motor behavior. New York, North Holland Publishing Co. Kugler, P. N., Kelso, J. A. S., and Turvey, &I.T. (in press) On the control and coordination of naturally devebping systems. In J. A. S. K&J and J. Clark (eds.j, Development of human motur skill. New York, John Wiley. Lee, D. (198@ Visuo-motor coordination in space-time. In G. Stelmach and J. Requin feds.), Tutorials in motor be&&or. Amsterdam. North-Holland Publishing Co. Mace, W. M. (1977) James Gibson’s strategy for perceiving: Ask not what’s inside your head, but what your headf inside of. In R. Shaw and J. Bransford (eds.), Perceiving, acting und knowing. Hillsdale, NJ, ErIbaum. Michaels, C. F., and Carello, C. (1981) Dtiect perception, New York, Prentice Hall. Morowitz, H. J. (1978) Foundations of bioenergetics, New York, Academic Press. Pattee, H. I-I.i 19741 Discrete and continuous processes in computers and brains. In M. Conrad, N.Guttinger and M. Dal Ciu (eds.), Lecture notes in bio-mathemattis 4: Physics and mathematics of the nervszIpsystzm New York, Sptinger-Verlag. Patten, B. 6. (i379) Environs: Relativistic elementary particles for ecology. Paper presented at Oak Ridge National Laboratory, Tennessee. Prigogine, 1. (1978) Time, structure and fhtctuations. Sci.. 202,777-785. Reed, E. S. and Jones, R. K. (1978) Gibson’s theory of perception: A case of hasty epistemologizing? Phil S&i, 45,519-530. Reed, E. S. ( 197Fr The ontological status of natural laws. Unpublished manuscript, Center for Research in Human Learning, University of Minnesota, Minneapolis. Runeson, S. (1977) On the possibility of ‘smart’ perceptual mechanisms. ZIctind J. Psychol., 18, 172-179. Se&wick, H. A. (1973) The visible ho&on: A potential source of visual information for the perception of size ancf distance, (Doctoral 4Issertation, C~~rnellUniversity, 1973). Dissertution Abstracts International. 34, 1301B-1302B. (i _iiversity Microflls No 73-22,530). shaw, R. E.. and Cutting, J. (1980) Clues from an ecological theory of event perception. In U. Bell@ and M. Studdert-Kennedy (eds.)&nedand spoken language:Biological constraints on linguistic form. Weinheim, Verlag Chemie. Shaw, R. E., and McIntyre, M. (1974) Algoristic foundations to cognitive psychology. In W. Weiner and D. PaIermo (eds.), Cognition und the symbolic processe% Hillsdale, NJ, Erlbaum. shaw, R., and Turvey, M. T. (1981) Coalitions as models for ecosystems: A realist perspective on per~eptwi organization. In M. Kubovy and J. Pomerantz (eda.), PerceptuuZorganiz&ion, Hiisdale, NJ, Erlbaum. Shaw, R, TUmY, M.T., and Mace, W. (In press) EcoIogieal psycl~ology: Theconsequenceof a commitment to reaIism. In W. Weimer art-l D. Palermo @Is.), Cognition and the symbolic processes‘ IL Iii&We, NJ, ErIbaum.
CogniGon: The view jtom ecological realism
32.1
Turvey, M. T., and Shaw, R. (1979) The primacy of perceiving: An ecological reformulation of perception for understanding memory. In LG. N&on (ed.), Perspecrives on memory research: Essays in honor of UppsaibUniversity’s500th anniversary,Hi&dale, NJ, Erlbaum, Turvey, M. T., Shaw, R. E., Reed, E. S., and Mace, W. M. (1981) Ecological laws of perceiving and acting: In reply to Fodor and Pylyshyn (1981). Cog., 9, 139-195. Whittaker, R. H., Levin, S.A., and Root, R. B. (1973) Niche, habitat, and ecotope. Amer. MS., ?07, 321-338. Yates, F. E(1980) Systems analysis of hormone action: Principles and strategies. In R. F. Coldberger (ed.), Biologicalreguktion and development. Vol.III: Hormoneaction. New York, Plenum Press.
Cognition, 10 (1981) 323-329 @ Elsevier Sequoia S-A., Lausanne - Printed in The Netherlands
323
Linguistic knowledge ad llanguageacquisition VIRGINIA
VAtJAN*
Columbia lh versity
This paper presents several hypotheses about knowledge and knowledge acquisition that are relevant to problems of language acquisition, and in terms of them assesses one aspect of the study of language acquisition and makes suggestions about future progress. Knowledge is a mental state, and may be explicit o?:tacit. Knowledge may be of, about, or that something. Knowledge is organized in terms of propositions whose elements are concepts. (The discussion excludes knowledge how, even though such knowledge plays an important role in language acquisition.) We have and acquire knowledge, rather than knowing and learning knowledge. Learning is one method whereby we acquire knowledge. Why is it incorrect to say that we know or learn knowledge? It can be more easily seen with other propositional attitudes. Take desire. Say that one could characterize a rule system for desire that was mentally represented and via which people determined their desires. (Whether such a system exists is immaterial.) One would not. say that one desires the rules: the rules allow one to derive desires; the rules are mentally represented; a person has the rules. Similarly for every propositional attitude: one can have the rules that constitute the attitudt, but one can not bear that attitude toward the rules. There is good reason to think that knowing is like every other prapositional attitude. Knowledge represents a relation between a person and something which can be known of, about, or that. Knowledge has a content which states what is known. Grammatical rules comprise one’s knowledge of language. One has the knowledge, and therefore also has the rules which CPAMute the knowledge. If, however, one knew the rules, there would hav ; to be another set of propositions comprising that knowledge. Whenever on: attributes knowledge one must also attribute a set of propositions of which the knowledge consists. If one knows a rule then that knowledge must be characterized, and it cannot be done by using the same content as the rule. Only if one is said to have knowledge (in the form of having rules) rather than knowing knowledge (in the form of knowing rules) can the problem be avoided. *I thankJ. J. Katz for a discussion of these issues. Reprint requests should be sent to V. Valian,Pwchology Department, Columbia University, New York, NY 10027, U. S. A.
324
VirginiaVulian
The word ‘learn’ operates similarly. Just as the expression ‘knowing knowledge is ill-formed, so is the expression ‘learning knowledge’. The same sorts of things that can be known can be learned. The state of having knowledge can arise either as part of initial mental structure, or as the consequence of acquiring knowledge. The controversial problems begin in characterizing how knowledge can be acquired. One hypothesis put forward here is that knowledge can only be acquired via learning or via derivation from prior knowledge. Another hypothes;, is that there are two methods of learning, one being hypothesis-testing, the other being what I shall call fitting into or elaborating’ an already existent knowledge structure. Other candidates can either be reduced to one of these methods or ruled out. A third hypothesis is that there are different types of propositional knowledge, and different sources may be linked with different types. Taken together, the hypotheses suggest that each phenomenon of language acquisition is likely to be a mixed one. Thea, interpreting them as if they were unitary can result in incorrect reiection of certain learning methods. The first hypothesis, that knowledge can only be acquired via learning or derivation from prior knowledge, contrasts with suggestions that knowledge can be acquired by taking a pill, re-wiring a brain, being hit on the head, or the like. It may also contrast with suggestions that knowledge can be ‘grown’ (Chomsky, 1980), depending on how that metaphor is interpreted. Nothing definitive can be said in favor of or against any of the positions, but there is an argument that shows the difficulty of being confident that there are no boundaries on possible methods of acquisition of knowledge. There are two preliminary considerations. First, all conceivable methods require change in a mental state, and all methods outside of learning or derivation, b&h of’ which are mental processes, are methods involving change in a physical brain state which must be assumed either to cause or to be the same as chinge : n a mental state. Second, not all changes in brainstatescorrespond tochanges in mental states, so Lidit proponents of being hit on the head as an acquisition method must show what kinds of head hits will work. Thus, even if one denied that knowledge was a mental state and claimed it was only a special kind of physical state, it would still be necessary to specify the conditions under which being hit on the head could induce this special kind of physical .state. Taking a pill ard neural re-wiring connote more precise and localized brain changes than concussion, but there is no difference in princ2plc. Physical methods are also problematic in that we know of no cases where purely physical alterations have resulted in knowledge acquisition, or anything that looks like knowledge acquisition.
Linguistic knowledge and language
An apparent parallel exists in that not all mental processes result in knowledge acquisition, and thus proponents of learning must provide theories showing what kinds of mental changes result in knowledge acquisition. There is a difference, however, since ‘learning’ is, by definition, a method of know]edge acquisition. The preliminary considerations to one side, here is the argument. One can theoretically mimic a process or state to any degree of precision, If there is a difference between knowing something and acting in such a way that is behaviorally indistinguishable from knowing something, then it must be demonstrated that a proposed method of acquisition delivers a knowledge state and not a state mimicking a knowledge state. Learning does so by defmition, SO that with learning the problem is to come up with a correct theory. Other proposed methods do not necessarily lead to knowledge. One alternative is to deny the difference between knowing and acting in a manner behaviorally indistinguishable from knowing, with the consequence that a chess computer, say, and a chess player have the same relation to the rules of chess.’ If one denies the difference, however, there seems no point in talking about knowledge. One may as well talk merely of mechanisms which govern behavior. An analogy can be made with ESP. When magicians show that by clever sleight of hand feats of telekinesis can be understood as trickery we do not say, ‘Oh, that’s what telekinesis is, clever sleight of hand’. Rather, we say, ‘There’s no such thing as telekinesis, it’s all clever sleight of hand’. Clever sleight of hand is not an exposition of telekenesis. Thus, if we want to retain the concept of knowledge we must also maintain a difference between having knowledge and mimicking having knowledge. Given the reality of that distmction, an appropriate concern about a proposed acquisition model is whether it can in. principle arrive at knowledge. Again, for a learning theory or a derivetion theory, the in-principle question does not arise; we know knowledge can be acquired by learning or by derivation from prior knowledge. The only qucstion is whether we have sbecified the method properly. There seem, then, grounds for provisionally accepting the hypothesis that only learning and derivation from prior knowledge will allow knowLedge acquisition, The second hypothesis, that th.cre are only two methods oflanguage learning, hypothesis-testing and fitting into a knowledge structure, cannot be ‘Chess rules are different from linguistic rules. One can know the rules of chess, unlike the rules of crammar, because chessrulesare like facts&out the language, Knowledge of the rules of chess is P trivial lonent of chess-playing: one does not derives moves from chess rules unless one does not know fow to play. A chess player has another set of rules, implicit ones which are more sin~ilarto linguistic rules,that determine moves; that set the player does not kno--abut has.
argued for here. Rather, the obvious methods which do not fit either model (e.g., conditioning, reinforcement, and imitation) are dismissed as already ruled cmt,2 and most remaining methods are taken as variants of hypothesistesting or knowledge structure elaboration. Is ‘organ growth’ (Chomsky, 1880) a possible except ion? Organ growth may or may not be a learning theory. If it is not, it is subject to the problems mentioned above. Chomsky seems to suggest it is not by likening the language system to the visual system. The analogy is not helpful, however, because it is not clear that any knowledge is involved in either the development of the visual system, or as a product of that development. To task a simple example, in recognizing an object, and in knowing it is a chair, the knowledge is not in the visual system. One has learned that a certain visual configuration represents a chair, and made a link between cognition and perception. Thus, the knowledge situation is sufficiently confused in perception that making analogies with it does not clarify the status of linguistic knowledge and its acquisition. Chomsky also, however, likens language learning to Peircean adbuction which is a hypothesis-testing model with a heavily constrained initial hypothesis space. If that is a correct characterization, then an organ growth model would seem to be a learning theory. All other language learning models appear to be variants of a hypothesis model, a knowledL:-structure elaboration model, or a composite. Therefore, it will be useful to briefly characterize suchmodels. Ahypothesis-testing model has five components; it is in how the components are specified that the theories differ. The first component states what is acquired, the second what is innate, the third what the content of the hypothlesesis, the fourth the role of experience. The fifth component delimits the constraints on the hypothesistesting mechanism (e.g., how many rules it can change at one time, whether it can store unanalyzed strings, and so on) and the procedures the mechanism uses in testing hypotheses. The first two components highly constrain the hypothesis space; expe@ence serves as confirming or disconfirming evidence. *Strategies (Bever. 1970) and operating principles (Slobin, lY.73) have not been offered as learning theories. As Cramer (1976) points out, strategies might explain how children behave when they do not umkstand a structure 5ut say nothing about how children learn a structure. Operating principles are based either on commonly-encountered properties oflanguages (e.g., ‘avoid exceptions’ reflects the fact that languages are system&c) or on commonlgrencountered features of learners(e.g., ‘avoid exceptions’ refkcts the fact that learners tend to form rules), or both. They are compatible with any type of learning theory. Piagetian principles have also not been offered as a language learning theory, although cognitive PhenomeIIa predicted by Piagetim theory often occur just prior to or concomitant with language pha rtomena, the signifkange of which is unclear.
Linguistic knowledge and !anguageacquisition
327
Erreich, Valian, 2nd Winzemer ( 1980) and
*
pothesis, that there are different types of propositional knowledge and different sources for the different types. We can roughly characterize two types: Type I includes laws and principles, and theories and explanations, and tends to be innate or acquired via hypothesis-testing; Type II includes single facts and systems of facts (like tax.onomies) and tends to be acquired by elaborating a knowledge structure. The basic outline of such a model is that the concepts in which the new knowledge will be encoded are already present; learning involves hooking up a new combination, or new sub-combinations, and wherever possible connecting the new fact with prior knowledge to make it less isclated. Many of the facts we acquire are ‘direct’ facts: they are presented in 2% manner highly si;!lilar to their ultimate mental represt;ntation. A newspaper article about the A,:ademy Awards, for example, states who won an award for best screenplay. We know about screenplays, awards, movies, and so on, -and can fit the new knowJedge into our knowledge structure-about Hollywood, sayquite easily. Otiler facts, however, such as word meanings, are typically not presented directly: to fit them in with a knowledge structure may first require a procedure of hypothesis testing, in order ‘to determine what the facts are. In language acquisition both Type I and Type II knowledge must be acquired: many individual specific facts must be learned (e.g., the declensions of irregular verbs), as well as broader language-specific patterns that will be mentally accounted for in terms of rules and principl.es. It is thus quite likely that some knowledge is acquired via both methods of learning, and some pia derivation from prior knowledge (as when a judgment can be made about, say, the ambiguity of a sentence never heard before). If it is the case that different kinds of knowledge are being acquired simultaneously, by the operation of at least three different methods of knowledge acquisition and revealed in a performance mechanism which is also developing, each resulting phenomenon of language acquisition is likely to be a mixed bag:, ‘Pinker (1981) seems ‘io offer a mixed model, part hypothesis-testing, part knowledge structure elab
328
Vitginia Valian
and riot susceptible to unitary interpretations. Attempts to force unitary interpret&ions can thus give misleading pictures of language acquisition. Salient superficial factors can appear to be the only factors. Usually, good performance requires knowledge of general rules plus specific facts, as in whquestion comprehension, which involves knowledge of **?h-movementand subject-aux inversion, and also of iniividual wh-words and individual verbs (see Winzemer, 198 I for a working-out of this example). If a unitary interpretation is sought, the salience of generalizations that hold for performance on specific words can override attention to the concurrent acquisition of rules. A second case occurs when different domains of knowledge are being acquired simultaneously. If the child is acquiring phonological, syntactic, semantic, and pragmatic knowledge simultaneously there will be numerous compatible descriptions of her behavior, all of which can be correct. If, however, the availability of several different descriptions is coupled with reductionism, then some descriptions will be incorrectly rejected. For example, syntactic and semantic descriptions are often available for the same phenomena (take Brown’s 1973 discussion of syntactic and semantic cumulative complexity), which is what one would expect if both syntactic and semantic knowled&?as:?being acquired simultaneously. But the tendency to see syntax as reducible to semantics results in syntactic descriptions being rejected if there is a semantic description available. (For further discussion see Valian, 1981.) The current situation in the study of language acquisition is one in which the theories make few predictions about phenomena, and the phenomena radicalIy underdetermine the theories. (Further, many of the theories are so non-specific that they cannot be evaluated on any grounds; notable exceptions include Hamburger, 1980, Wexler and Culicover, 1980, Pinker, 198 1, Erreich, et al., 1983.) How can the theories and phenomena come in closer wntact? The first suggestion is, paradoxicaIly, that theorists temporarily put aside attempts to account for most current phenonema, and concentrate on deriving predictions from their theories about new phenomena. The problem with most known phenomena is that, being mixed, they frustrate efforts at theory construction. An example is telegraphic speech, for which we tend to look for a single account. However, some word omissions may be due to one cause, while others have a different cause. Valian (1981) shows that the frequent absence of determiners in children’s speech can be explained by the optionality of determiners within noun phrases. But irheabsence of be as auxiliary must have a different explanation. Thus, ‘telegraphic speech’ may be une name covering several different processes, with one process linked to the acquisition of categories. Telegraphic speech, as a phenomenon, cannot helpfully constrain
Linguistic knowledge and languageacquisition
329
learning theories if it is not actually a single phenomenon. Eventually the old phenomena will have to be explained or convincingly demonstated to be outside the domain of language acquisition, but we will be in a better position to do that with stronger theories. The second suggestion is directed to the need for phenomena that will bear more Directly on learning theories, by being less mixed. There might be two ways of arriving at purer phenomena. One is to avoid clearly mixed phenomena unless guided by a theory which analyzes the types of knowledge and the methods of acquisition invojved, or unless it is possible to control for all but one knowledge type and one acquisition method. An example of a clearly mixed phenomena is increasing MLU with age, which reflects all types of knowiedge acquired in all ways. Thus, despite its robustness, the phenomenon is not Woretically useful. Another way of arriving at purer phenomena is by successive approximations: one can analyze a phenomenon according to its probable knowledge types and acquisition methods, using the results to suggest future observations that will be localized to a particular knowledge and acquisition type. The phenomena will thereby be more directly relevant to the construction and confirmation of theories. References Anderson, J. R. (1980) Cognitive Psychology and its implications. San Francisco, W. H. Freeman. Bever, T. G. (1970) The cognitive basis for linguistic structures. In J. R. Hayes (ed.), Cognition and the Development of Language. New York, Wiley. Brown, R. (1973) A Flist Language. Cambridge, HarvardUniversity Press. Chomsky, N. (1980) Rules and Representations. New York, Columbia University Press. Cramer, R. F. (1976) Developmental strategies for language. In V. Hamilton and M. 13.Vernon (eds.), It?teDevelopment of CognitiveRecesses. New York, Academic Press. Erreich, A., Valian, V., and Winzemer, J. (1980) Aspects of a theory of language acquisition.J. Child Lu?ag.,7,157-179. Hamburger, H. (19gOj A deletion ahead of its time. Cog,, 8, 389-416. Mayer, J. W., Erreich, A., and Valian, V. (1978) Transformations, basic operations and language acquisition. Cbg,,’6, 1-13. Pinker, S. (1981) A theory of the acquisition of lexical-interpretive grammars. In J. Bresnan (ea.), 7%e Mental Representatbn of GrammaticalRelatbns. Cambridge, MIT Press. Slobin, D. 1. (1973) Cognitive prerequisites for the development of grammar. In C. A. Ferguson and D. Slobin (wk.), S&dies Qliki LanguageDevelopment. New York, Holt, Rinehart and Winston. Valian, V. (1981) Syntactic categories in the speech of young children. Unpublished manuscript. New York, Columbia University. V&n, V., Winzemer, J., and Erreich, A. (1981) A ‘Iittle linguist’ model of syntax learning. In S. Tavakolian (ed.), Law AcquWion and L&uistic Theory. Cambridge, MIT Press. Wexler, K., and Cuhcover, P. (1980) Formal l+inciples of LunguageAcquisition. Cambridge,MIT Press. Winxemer, J. (1981) A lexical-expectation model for children’s comprehension of whqil.%tions. Unpublished Ph.D. dissertation. New York, CUNY Graduate Center.
of
Cbmition, 10 (1981) 331-336 @Elsevier Sequoia S.A., Lausanne - PGnted in The Netherlands
331
Human memory and the information-processingmetaphor MICHAEL J. WATKINS* Rice University
Cognition: Au International Journal of Cognitive Psychology is healthy at ten. It was born in a veritable explosion of cognitive research, research that for the most part was cast within an information-processing framework. As the pages of this and other cognitive journals testify, this framework has fared extraordinarily well over the ensuing years, not least in my own area of: research, the psychology of memory. Indeed, it appears that research not conceptualized in information-processing terms is all but extinct. That much of the research from our laboratory provides something of an exception would seem to call for comment, and such comment forms the substance of this statement. One of the weaknesses of an information-processing approach to the study of memory is that it is so successful. It is hard to think of a finding it cannot accommodate. Select any experimental finding from the memory literature and any information processor worth his salt will have no trouble in putting together a combination of mental processes and structures to serve as a theory for it. Moreover, an information-processing theory is not only easy to create but, once created, any encounter it may have with data is unlikely to prove fatal, It might seem that we should leave such a happy state of affairs well alone, but I suspect there may be room for skepticism. A key factor in the survival of information-processing theories is the odd way they have of proving to be more subtle than they at first appear. All too often it turns out that the critic of a theory has overlooked an implicit assumption or some other subtlety. Also, in the rare case where a theory is conceded to be at variance with the data, the problem is usually resolved with only minor tuning. Another reason for the immunity of an informationprocessing theory to effective criticism is that it is unlikely to attract much attention. This is because the information-processing era has brought with it the luxury of personal theories, so that researchers reserve most of their attention for their own, special theory. The upshot of this state of affairs is that theories typically survive as long as, but no longer than, the active interest of their creators. It can, of course, be questioned whether the proliferation of personal theories is really undesirable, for, after alI, research still gets done. A perspec*Reprint requests should be sent to: M. J. Watkins, Department of Psychology, Rice University, Houston, Texas 77001, U.S.A.
332
Michae,~J. Watkins
tive on this question can be gained by imagining the consequences of imposing sever12 restrictions on the amount of theorizing allowed in research reports. Natlurally, authors would still be responsible for reviewing relevant previous research, but suppose they were permitted on’ry minimal space to interpret, explain, rationalize, or otherwise discuss their findings. It seems to me that the !::onsequences might not be unduly harmful. I find it curious that we resi?arch psychologists often have such a hard time understanding the theory in each other’s articles yet feel competent to evaluate the reported research on the basis of the -method and results alone. The question arises, therefore, of whether the benefits of contemporary theorizing are worth the efforts required to read it or indeed the cost of publishing it. Moreover, there is a fur&r cost of persona.-1 theories that, though perhaps less obvious, is probably more important. Testing the predictions of personal theories usually involves complex designs and a concern with high-order interactions. In other words, personal theories lead to personal research, which means that when the interest of the theorist ebbs, not only his theory but also his rearch is in danger of being left high and dry. Over the past few years it has increasingly become our practice to report research with as littfe theorizing as editors will let us get away with. We now tend to avoid using reghtration, encoding, storage, retrieval, and other information-proce-csing terms. Clearly, we face the question of whether we are overreacting. In particular, is there anything to be gained by totally excluding the information-processing language rather than using it sparingly? I beIieve that perhaps there is, that it may be important to see just how far it is possible to get aloh.g with a COK;’ +te abstention from information-processing constructs This belief rest;, m part on a suspicion that there may be some fundamental weaknesses in the conceptual underpinnings of the entire information-processing approach. UnderEying the shortcomings of information-processing theorizing is its adoption of the spatial mode1.l According to this model, memory involves three distinct stages. First, information is put ‘into’ memory; then for some period of time this information is retained ‘in” memory; fmally, recollection O~CWSwhen information is taken ‘out of’ memory. This spatial conception of memory is, of course, by no means an invention of the information processors, for it was at least imp&d in earlier theories. Nevertheless, information processors have based their theorizing more squarely on the model than have most previous theorists, drawing a sharp distinction between its stages 1
spatial model of memory has recently been discussed by Roediger (1980), though it sh.ould be mHed that I use the term in a brwder sense than Woedigerdoes, ;io that I would call spatial some of the metapbxs (such as the lock-and4cey and the resonance metaph,hors) that he c& non-spatial.
Human memoy and the information-processing metaphor
333
and assigning to each an explanatory role. Indeed, with the informationprocessing approach to memory the spatial model has become the foundation of a gigantic theoretical edifice, and as such it would seem prudent to inspect it for possible structural flaws. One potential flaw in the spatial model concerns its elaborateness relative to that of the basic memory paradigm. The basic memory paradigm involves simpl>r one or more ‘treatments’ or subjects, which typically include the presentation of to-be-remembered material in the form of a study list, followed by the measurement of recall urz3er one or more test conditions. Typically, the findings submit to more than one inteTretation within the framework of the spatial model: This is true not only for the elaborate versions of tht: model given in personal theories, but even for the model in its raw unelabclm rated form. I have discussed this point elsewhere (Watkins, 1978), and it L enough for now to illustrate it with an example. Suppose that two groups of subjects are presented with, and then tested for recall of a word list, and that at all phases of the procedure the two groups are treated identically with the single exception that for one group the words have a high frequency of everyday usage and for the other they are comparatively rare. Suppose further that, as seems likely, the group given the more common words tends to recall more. This frequency effect could be attributed to the input stage of meTory, in that encoding could be richer or deeper for common words, Alternatively, the effect could ar..se at the retention stage, in that during retention subjects in the high-frequency condition are likely to encounter more stimuli or think of more ideas that are associatively related to the to-be-remembered words, and such encounters or thoughts could strengthen the memory traces. 14nd of course the frequency effect could also be attributed to the output stage, as with the assumption that common words a-e more easily located within the memory system. Although such interpretations may d.iffer in their plausibility, it may not be possible to design an experiment that would fbrmally distinguish between them. If so, even the basic ve.rsian of the spatial model would seem to have a superfluity of explanatory power. Such problems notwithstanding, the three-stage model of memory is SO well ingrained that the possibility of doing without it may not be entirely obvious. To illustrate this issue, consider the concept of storage and its status in memory theorizing. More specifically, consider the argument that if we witness an episode and can later recall it, then surely we must in some small way have been changed by that episode, we must have had stored within US information pertaining to it. With.out wishing 110challenge the reasonableness of this argument, I am of the view that whether it is essential for our theories to incorporate the idea of storage depends upon the function they
334
MichaelJ. Watkiprs
are intended to serve. If they deal with the physiological substrate of memory then it is indeed hard to see how they could avoid including reference to storage, though it is of interest to note that the physiological psychologists who use such theories think in terms of procedures for examining storage directly. For most other experimental psychologists, however, the purpose of a theory of memory is to bring order to data concerning the relation between what the ej;perimenter does to the subject and the subject’s subsequent behavior, and here reference to storage would seem to be optional. Much of the research from our laboratory can be seen as relevant in one way or another to the evaluation both of this option in particular and of the three-stage model in general. For instance, in several studies of recognition memory we have shown in a variety of ways that recognition performance v&ties with test condition, especially with the context in which the test item is presented (e.g., Todres and Watkins, 1981; Watkins, 1974; Watkins, Ho, and Tulving, 1976; Watkins and Tulving, 1975). Contrary to a not uncommon assumption, these findings, along with those of studies from many other laboratories (see, e.g., Watkins and Gardiner, 1979), demonstrate that as a test of whether an item is ‘available in storage’, the recognition procedure is no more adequate thaa any other we know about. Given, then, the difficulty in operatioiralizing even the essential components of the three-stage model of memory, we hat/e spent considerable energy exploring the option of rejecting this model, and therefore the entire information-processing approach, in favor af a simple, functional approach. This exploration has involved (1) gathering experimental evidence on some of the more obvious questions that arise when memory is considered in the context of t5r,ebasic memory paradigm; (2) considering how, in the absence of the mechanisms of the informationprocessing approach, the facts of memory might be explained; and (3) seeing whether a :functional approack to memory can be usefully applied to phenomena that appear particularly well suited to an infocmation-processing interpretation. I will briefly sketch these three endeavors irr ,tum. From the standpoint of the basic memory paradigm, it would seem to be important to know about how item recall is affected by the conditions under which the items are both studied and tested, Of the two, we probably know less about the effects of test conditions, ani su it is here that, to date at least, we have concentrated our major efforts.2 For the most part, these efforts have consisted in determining the relative effects of various types of cues and the contingency relations among them (e.g., Tulving and Watkins, 1975 ; WatzThk is not to deny, of course, that any conclusions about the effects of test conditions have to be quaWkd with aspect to study conditions (see Tulving and Thomson, 1973).
Human memory and the information-prowsing meinphor
335
kins and Todres, 1978). It should be noted that I am using ‘cues’ very broadly, broadly enough to include the free recall instruction and the test item of the recognition procedure (see Watkins, 1979). Determining the effectiveness of varic us study and test conditions does not, of itself, provide a means of explaining memory, of accounting for why some things are remembered and others forgotten. Indeed, if we are to forgo the mechanistic, information-processing explanations, it is not even clear what form explanation should take. This is too big an issue to address here, but for our part we have sought explanation in terms of laws or principles, by referring the particular to the general. A generalization we have found especially useful is one we refer to as the cue-overload principle, according to which the probability of a cue effecticg recall of a given item varies inversely with the number of items the cue subsumes.” This principle can be invoked as a ready interpretation of the list-length effect, the advantage of using categorically structured lists, subjective organization, the effects of extralist cuing, and various paired-associate findings (see Watkins, 1979). Moreove; we have applied it to, and confirmed its predictions concerning, the buildupand-release-from-proactive-inhibition effect (Watkins and Watkins, 1975) and the inhibitory effect of part-set cuing (Mueller and Watkins, 1977; Todres and Watkins, 1981; Watkins, 1975). I should add that we have also used the cue-overload principle to develop a retroaction procedure for exploring the functional relation between cues (Watkins and Watkins, 1976). Some idea of how widely our relatively atheoretical approach to memory may be usefully applied can be gained by seeing how well it copes with topics that seem well suited to an information-processing approach. For some time now we have been studying two topics, sensory memory and rehearsal, that regularly furnish textbook illustrations of this approach (Craik and Watkins, 1973; Watkins and Graef’e, 198 1; Watkins and Todres, 1980; Watkins, Watkins, and Crowder, 1974; Watkins and Watkins, 1980a, B). In reporting this research we have, over the years, gradually reduced the amount of reference made to information-processing constructs, and I now believe that, as in our more recent reports, all $uch references could Ive been omitted without obvious loss. Although the future is something l prefer to let take care of itself, I can ski that for the present we are continuing with most of these areas of research. Just how far we can get with our functional approach to memory remains an open question, but at the very least, the mere appreciation of its 3 It is perhaps appropriate to suggest that the cue+verload principle should not be regardedas a personal theoty. Not only is it a simple idea, but it constitutes the common ground to a variety of more specific ideas put forward by a large number of researchers in the contexts of a variety of relatively specific domains.
336
MichaelI. WutiEas
limitations should help clarify the advantages gained from theorizing in elaborate information-processing terms. Indeed, just as fish may have a hard time coming to grips with the cloncept of water, so may we lose our grasp of the information-processing metaphor under the sheer vastness of its popularity.
Refermces F. 1.
Watkins,
The role
verb. Behav.,12.599-607. Mueller, C. W., and Watkins, M. J. (1977) Inhibition from cuing with recall targets: A cuesverload intezpretation. (. verb. Learn.verb. Behav.,I6,699-709. Roediger, H. L (1980) Memory metaphors in cognitive psychology. Mem. Cog., 8, 231-246. Todres, A. K., and Watkins, M. J. (In press) A part-set cuing effect in recognition memory. J. exper.
Psychd.: Hum. km. TuSv&
Mem.
E., and Thomson, D. M. (1973) Encoding specificity and retrieval processes in episodic memory. Aychol. Rev., 80,352-373. Tubing, E., and Watkins, M. 1. (1975) Structure of memory traces. Psychol. Rev., 82, 261-275. Watkins, M. 3. (1974) When is recall spectacularly higher than recognition? J. exper. Psychol, 102, 161 -163. Watkins, M. 3. r.h975) Inhibition in recall with extralist ‘cues’. J. verb. term. verb..Behuv.,I#, 294303. Watkins, M. J. (1978) Theoretical issues. In M. M. Gruneberg md P. E. Morris teds.), Aspects of humanmemury. London, Methuen. Watkins, M. J. (1979) Engrams as cuegrams and forgetting as cue overload: A cueing approach to the structure of memory. Jn C!. R. Puff ted.), Z&e stnrciure of memory. New York, Academic Press. Watkins, M. J., and Gardh%er, J. M (1979) An appreciation of generate-recognize theory of recall. J. ve& bn. verb. Behav., 18,687-704. Watkins, M. i., md *Zraefe, T. M. (In press) Delayed rehearsal of pictures. J. verb. Ileum. verb. Behov. Watkins, V 3., %, Z., and Tulving, E. (1976) Context effects in recognition memory for faces. J. ve& km. verb. Behav., i&505-517. Watkins, M. J., and Todres, A. K. (1978) On the relation between recall and recognition. J, verb. L&uri.wrb. Behav., 17.621-633. Watkins, M. J., and Todres, A. K. (1980) Sufhx effects manifest and concealed: Further evidence for a 2O-seco& echo.4 verb. Learn..verb. B&tntv..19.46-53. Watk& %f.+ J., and Tulving, E. (1975) Episo&c memory: When recognition fails. J. exper. Rychol.: Gen.. IW, 5-29. Watkins, M. J., and Watkins, 0. C. (1976) cue-overload theory and the method of interpolated attributes. BuQ plyahort, &x.. 7.289-291. Watk& If. J., Watkins, 0. C, and fiowder, R G. (1974) The modality effect in free and serial recall as a function of phonological similarity. 1. verb. lkwm. verb. Behav., 13,430-447. Watch% 0. C, and Watkins, M. J. (1975) Buildup of proactti inhibition as a cueoverload effect. 1 cxpa. &ycM.: Burr. Leant. Mem.. I, 442-452. 0. C, and Watch M. J. (1980~) Echoic memory and voice quality. Mem. Qg., S,, 26-30. 0. C., and Watkins, M. J. (198Ob) The mod&y (effect and echoic persistence. J. exper. Agcird.: Gen., 109.251-278.
Cognition, 10 (1981) 337-340 @ Elsetier Sequoia &A., Lawanne - Printed in The Netherlands
337
A position note on natural language understanding and artificial intelligence YQRICKVJILKS” University
of Essex
What .follows is in no sense a ,logical questions I am agnostic, though admitting that, whatever dislalmers their promake/ all artificial intelligence (AI) in fact incorporate example, even th3se at what one might call the “engineering” spectrum of AI research disclaim all theory-tend nonetheless to construct English, say, that from left-to-right, suggest that rightleft processing certain constructions explain such a preference. towards that end of the spectrum myself, opposite which I tend to identify systems independent of their instantiation in machines, humans or animaZs. This is an AI version of linguistic competence theory, and no m3re appealing for that: it is liable to miss the distinctively AI insights that come precisely from consideration of processing constraints. But, I repeat, whatever an .\I worker’s chosen place on that spectrum, doing psychology is not his job. Whether one calls it division of academic la bour, or mere trades unionist protectionism, I believe psychological speculation and testing is best left in the hands of psychologists. Some fifteen years ah, I, began to publish papers on programs to “parse English text semantically”, an enterprise that has been widely misunderstood since then: as, for example, claiming that such a program could not take account of the (syntactic) fact that, say, English determiners tend to occur to the left of adjectives. This was clearly a misunderstanding; no more was intended by the original claim than that the work of parsing English can be done via a structure that is plausibly semantic in nature, with no autonomous syntactic component. This claim still seems to me plausible and proof against Katnian knock-down answers in terms of arms and weapons having “the same meaning” while be;ng associated with different syntactic features. ---
*Reprint requests should be sent to Y. Wilks, Department of Language 8s Linguistics, University of Essex, Wivenhoe Park, Cokhester CQ4 3 SQ, England.
338
Y. Wilks
I feel under no obligation to agree that they “have the same meaning” and, moreaver,, it now seems to me that the recent emphasis on “perspectives” in AI 6.a~in Bobrow and Winograd’s HRL language) can offer a formalism in wlmichto make such a position concrete: the knowledge and meaning structures of armr and weapons could plausibly differ if only one of them could “seen from the perspective”’ of countability. Let me restate the principles behind that fifteen-year old approach to “semantic parsing”: (a) that language is fundamentally a linear, segmentable, phenomenon, even at the semantic level, rather thau a hierarchical one; (b) semantic and knowledge-structure dictionary entries are the fundamental data structures for parsing, but there are items in the system (“meaning skeletons”’ for phrases and clauses) that are not reducible to dictionary entries: I called these temp2ates. In brief, this approach does not accept a simple version of the Fregean “principle of compositionality”. (c) there is a very general algorithm for selecting fillers for slots, and hencp overall structures of sentences and texts: I called it “preference”. It was based on a notion of semantic coherence-in short, accepting the semantically densest reading-and never rejected readings, only preferred some to others. (It was, in a clear sense, the opposite of what is known as “constraint analysis”: that alttznatives are counted out not in). These principles still seem to be broadly correct. and I take encouragement not only from, say, recent work on the key role semantically coherent noun groups play in garden path sentences and hence in parsing generally, but also from the general drift of linguistic theory in the last ten years. By that I mean the move towards a more surface-orientated syntax, concerned with th.e role of dictionary entries and slot fiig rather than transformations. If it be replied that one who argues for “semantic parsing” can hardly be cheered by a move towards a surface syntax, I demur. Semantic parsing of the type I intend is deep& superficial. until extended by inference structures and knowledge bases, and I would maintain that even against those who have done the same kind of parsing as that I advocate and called it “conceptual” and “deep”. I have many times pointed out in print its relentlessly, and in my view quite correct, superficial properties. The key question will be the ability of those now engaged in syntactic parsing, such as Marcus, to produce interesting and significant generalizations not
reducible to semantic ones. We are hovering near an old problem: th;tt of semantic primitives. I think it is now clear that there are a lot of b.ad ways of defending such entities: as part of an innate brain language, for example. That does not mean that there are not good ways, and I think one of those: is to be found near the phrase “I%oceduraISemantics”. I have attacked a number of expositions of this no-
Natumllanguageimierstanding and art@kkl inte&ence
339
tion, but I think it still offers a potential coup to AI: the construction of a distinctive theory of meaning (which would, ambulando, give procedural meaning to such items as semantic primitives) which was formal and defensible while not being, at bottom, reducible to model theoretic semantics, verificationism, or the banality of brain (or machine) hardware. Another good AI idea that still has much to offer, I believe, is frames, or scripts. I have never accepted that these still insufficiently defined items can guide parsing or be proto-text grammars. The evidence is too strong that we can do what is needed, in initial parsing at least, with weaker structures. Where I do believe them essential, however, is in the stage that immediately succeeds initial parsing. Only with their aid, I have argued, can we make sense of the very simplest utterances that break what I would call preference restrictions. Many such phenomena would be called syntactic by others, but no matter. On my view John ran a mile breaks the pre-reference of run for no object, and is, in that sense, preference breaking or metaphoric. Even such trivial cases can, I believe, be profitably subsumed under a general pattern-matching algorithm (forwhich I have specified a sample version) %iiatmatches preferencebreaking items (the above as much as “my car dzinks gasoline”) against fi-amelike structures to determine an interpretation via what is normally the case for the mentioned entities: gas is normally used by a car (so that interpretation is substituted in a text representation of the sentence), just as running normally extends a distance such as a mile (with corresponding effects on the representation). This sort of approach is, I believe, the appropriate use of frame-like stru= tures (incorporating detailed factual knowledge about the world) in the oralysis of sentences and texts, given a very general assumption that language is inherently, not incidentahy, metaphoric or boundary breaking, and that human comprehension is mapped by a general knowledge-based procedure imposing top-down coherence on what we read and hear. One final area of research, now booming in AI, should be mentioned: the analysis of conversation in terms of plan structures, beliefs and perspectives, loosely what has been known in philosophy and linguistics as speech acts. “Speech acts are dead” said an eminent linguist to me the other da;, which I would interpret bysayingthat philosophyandlinguistics encountered problems with the notion that their theoretical machineries did not allow them to solve. In the case of philosophy, the assumption of an irreducible and knowableby-others notion of intention seems intractable for any procedural account. In the case of linguistics, the machinery of generative grammar made it impossible to bring belief structures (particularly of different individuals, including their beliefs about each other) to bear on the “analysis of utterances.
I believe AI work in this area will be very fruitful and am actively concerned with it myself: partly because the notion of plans can assimilate speech to non-linguistic action in the $vayAustin originally intended, and partly because the perlocutionary effect-the goal of the act of uttering-can be kept firmly in view in an AI account, whereas it tends to get lost in accounts where the goal of speaking seems only to be understood, rather than to achieve some concrete end. I do not believe that such AI work will justify any particular theory of speech acts; on the contrary, I anticipate that the existing philosophical distinctions and tern.inology will disappear, or, at best survive as primitives of system organization. There might well be a group of rules clustered under the label THREAT or PROMISE in a useful system of conversational analysis of t!;e future or, on the other hand, the rule taxonomy used might have no commonsense interpretation. I hope devoutly for the former outcome, bc t we shall see. One danger I see to theoretically interesting natural language analysis in AI is the trend to expert systems: on that view, language about car repair or electrical circuits, say, becomes no more than a side-effect (the word is much used) in a system that plans such activities. This would be a parody world of Wittgensteitian linguistics, and some AI workers, seeing the effect of this trendsatisfying as it may be to commercial interests and Government sponsors -may soon come flocking back to the shelter afforded by the skirts of “generalized competence” and an “innate language faculty”! i ’
Cognition
341
Cumulative Author Index of Volumes The volunte numbers are in boldface type and followed by the page rtumber. Ades, Tony, lo,7 Alegria, Jesus, 7,323 Allen, Rhianon, 6,189; 8,175 Ammon, Mary Sue, 7,3 Amsel, Eric, 7,99 Anderson, John R., 8, ‘73 Antinucci, Francesco, 7, 145 Baars, Bernard J., 4,177 Bacharach, Verne R., 4 281 Bpddeley, Alan, IO, 17 Baron, Jonathan, 2,299 Beardsley, William, 6, 117 Bellugi, Ursula, 1, 173; 3,93; 4,45 Bertelson, Paul, 7,323 Bevel, Thomas G., 3,83 Biederman, Irving, 7,285 Binks, Martin G., 5,47 Blums!ein, Sheila E., lo,25 Bonvuhan, John D., 2,435 Boons, Jean-Paul, 2,183 Bornstein, Marc H., 6,89 Bower, Thomas G. R., 1, 47,165; 3,29 Brainerd, Charles J., 2,349 B$ansford, John D., 1,211 Br,egman,Albert S., lo,33 Bresnan, Joan, lo,39 Broadbent, Donald E., 10, 53 Bronckart, J. P., 2,1Q7 Brown, Roger, 4,125; $73,185 Bruner, J. S., 3,254; Bucci, Wilma, 6,55 Byrne, Richard, 5,287 .
Caplsn, David, 2,269 J 10,59
Caramazza, Alfonso, 3,227; 6,117;9,117 Carello, Claudia, 10,3 13 Carr, Thomas H., 4,28 1; 9,73
Cary ,
Luz, 7,323 Chomsky, Noam, 1 , 1 1,407 Clark, Eve V., 2,161 Clark, Herbert H., 8, 111; 9,311 Cohen, Gillian, 9,59 Cohen, L. Jonathan, 7,385; 8,89 Cooper, William E., 6, 135 Croft, Karen. 8.369 Cutler, Anne, 7,49; lo,65 Cutting, James E., lo,71 Delis, Dean, 5, 119 Deutsch, Werner, 6,155 Dickson, W. Patrick, 5,215 Donaldson, Margaret, 3,341 Dore, John, 2,45 1 Dover, Arlene, 7,99 Dresher, B. Elan, 4,32 1; $147,377
Duranti, Allessandro, 7, 145 Edwards, Derek, 2,395 Ehrlichman, Howard, 4,3 1 Eimas, Peter D., lo,79 Ellis, Andrew W., 7,413 Erdelyi, Matthew Hugh, 4, 311;9,23 Erreich, Anne, 6,1; 7,317 Evans, J. St. B. T., 1,373; 3,141,387; 5,265
Fairweather, Hugh, 4,23 1 F’auconnier, Gilles, LO,85 Finkelstein, Shira, 4,3 11 Fischer, Susan, 1,173 Fischler, Ira, lo,89 Flavell, John H., 8,369 Fodor, Janet Dean, 6,291; 8,417 Fodor, Jerry A., 1,83; 6,229; ‘7,49,93; 8,263; 9,139 Ford, Marylin, 6,35 Forster. Kenneth I., 2,3 19 Franks, Jeffery J., 1,211 Frazier, Lyn, 6,291; 8,417 Freeman, N. H., 8,243 Frenk, Samy G., 1,97 Garrett, Merrill F., 1,359; 8,263; lo,97 Garvey, Catherine, 3,227 Gebert, Lucyna, 7,145 German, Rachel, 4,189 Glass, Arnold L., 3,3 13 Gleitman, Henry, 1,137 Gleitman, Lila R., 1,137; 10,103
Goldin-Meadow, Susan, 4, 189 Goldstein, Louis M., 2, 279 Goodluck, Helen, 7,85 Green!, Bert, 9, 117 Grieve, Robert, 5,235 Grober, Ellen H., 6,117 Grosjean, Francois, 5,lO 1 Gross, Charles G., 6,89 Hakuta, Kenji, 9,197 Hamburger, Henry, I), 389 Hamlyn, D. W., 10,115
CumulativeAuthor Index
342 Hampson, P. J., 6,79 Hatano, Giyoo, S,47 Healy, Alice F., 10,119 Herrell Nadeanne 4,3 ! 1 Hermsiein, 8. J., i, 301, 419 Hi&, William, 4,215 Hobbs, S. B., 6,15 Hochberg, Julian, 10,127 Holmes, Virginia M., 6,35; 7,363 Holyoak, Keith J., 3,313 Hoogenraad , Robert, 5, 235 Hopkins, J. Roy, 2,385 Hornstein,Norbert,4,32:, 5,147,377
Horton, Marjorie S., 8, 227 Hughes, Jennifer, 3,41 Hupcey, John A., 3,307 Ibbotson, N. R., 9,125 fnhelder, Barbel, 3, 195 hwin, David E., 10, lit5 Jacobson, Sandra, 2,385 Je?nnerod, Marc, 53; 13,135 Jensen, Arthur R., I, 427 Johnson- Laird, Philip N., 1,s:; 5,189,& i<“*; 10, : 39 Jonides, John, 10,145 Jorm, Anthony F., 7,19, 421 Kagan, Jerome, 2,385 Kahneman, Daniel, 7,409 Kalish-Landon , Nancy : 2, 451 Karmiloff-Smith, Annette, 3,195; 10,151 Kean, Mary-Louise, $9; 7,69 Keil, Frank, 10,159 Kemper, Susan, 9,385 Kessen,William, 10,167 Kimball, John, 2,15 Y&a, Edwmi S., 3,93; 4,45
Klosek, John, 7,61 Kolk, Herman H, J., 6, 353 Kosslyn, Stephen Michael, 10,173 Kulik, James, S, 73 Lackner, James R., 1,210, 359; 2,279; 4,3Q3
Lane, Harlan, 3,351; 5,101 Lang, Stephen, 3,359 Langendoen, D. Terence, 2,451 Langer, Jonas, 1,105 ; 3,9; IO, 181 Langford, I., 7,363 Layzer, David, 1,265, ?3,454 Leontiev,A.N.,1,311 Levelt, Willem J. M., 10, 187 Lewis, Selma, 5,333 Lieberman, Philip, 2,59 Lindb”lom,Bjom, 4,99 Lloyd, S., 8,243 Locke, John L., 6,175 Loftus, Elizabeth F,, 110, 193 Longuet-Higgins, H. Christopher, IO, 197 .Lugg,Andrew, 3,289 Xuria,A.R.,1,311;3, 377 MCCloskey,Michael, 9, 117 Mace, W. M., 9,237 McGarrigle, James, 3,341 McLanahac, Alexander G., 8,227 McNeill, David, 10; 20 1 Markman, Ellen, 3,213; 8,227 Marshall, John C., 10,205 Marslen-Wilson,William, 831 Maturana, Humbcrto R., I,97 Mayer, Judith Winzemer, 6,1;7,317
Mehler, Jacques, 3,83 Miller Bruce, 4,3 11 Miller, Geo,ge A., 10,2 I 5 Miller, R., 2,49 Miyake, Naomi, 5,215 Miyake, Yoshio, 5) 47 Moore, Timothy E., 7,285 Morais, Jod, 3, 127; 7, 323 Morris, P. E., 6,79 Morton, John, 4,309; 7,35; 9,125; 10,227 Motley, Michael T., 4, 177 Mounoud, Pierre, 3,29 Murray, Diarmid, 5,235 Muto, Takashi, 5,215 Navon, David, 6,223 Neisser, Ulric, 3,307; 4,215; 6,169; 9,1 Nelson, Katherine, 4,13 Nelson pKeith E., 2,435 Newcombe, Freda, 10, 209 Newstead, S. E., 5,265 Ninio, Anat, 7,125 Norman, Donald A., 10, 235 Olbrei, llmar 2,3 19 Osherson, Daniel N., 3, 213; 4,203; 6,263; 9,35; 10,241 Paccia, Jeanne M., 6,135 Palermo, David S., 3,245 Parkes, C. H., 8,263 Paterso.?, J. G., 1,47 Pechmsnn) Thomas, 6, 155 Perfetti, C. A., 2,95 Peters, Ann M., 8,187 Petitto, Laura A., I, 177 Petrey, Sandy, $57 Phillips, W. A., 6, II 5 Pierson, Lee, 3,359 Pinker, Steven, 7,21?; 10,243 Pisoni, David B., 10,249 Pasner, Michael I., 10,261 P,att, F. R., 6,15
Cognition
Premack, Davrd, El,25 1; 7,333 Putnam, Hilary , %,13 1; 3,295 Pylyshyn, Zenon W., 3, 57; 9,139; 10,267 Reber, Arthur S., 5,333; 6,189;8,175 Reed, E. S., 9,237 Rips, Lance J., n, 145 Rouinson, E. J., :.q,363 Robinson, W. P., $, 363 Rose, Hilary, 2,479 Rose, Steven P. R., 2,479 Rumelhart, David E., IO, 235 Savin, Harris B., 2, 147, 213: 257 Schaerlaekens, A., 2,371 Schane, Sanford A., $35 1 Schank, Roger C., 5,133 Schiinbach, Peter, 5,181 Schnaitter, Roger, 3,79 Schunk, Dale H., 8,111; 9,311 Schwartz? Stephen P., 7,301 Seeidenberg,Mark S., 7,177 Seligman, Martin E. P., 4, 189 Shallice, T., 1,335 Shaver, Phillip, 3,359 Shaw, R. E., 9,237
Shipley, Elizabeth F., 1, 137 Shipstead, Susz-t G., 8,369 Shultz, Thomas R., 7,99 Sinclair, H., 1,317; 2,107 Sinha, C. G., 8,243 Siple, Patricia, 3,93 Slater, Anne Saxon, 5,119 Slobin, Dan 1.,7,3; 10,275 Smith, Edward E., 9,35 Smothergill, Daniel W., 5,251 Solan, Lawrence, 7,85 Sorensen, John hf., 6,135 Spelke, Elizabeth, 4,215 Sperber, Dan, 10,281 Staddon, John E. R., 10, 287 Stein, Judy B., 9,23 Stevens, Kenneth N., 10, 25 Strauss, Sidney, I, 105, 329; 3,155; 10,29.$ Studdert-Kennedy, Michael, 8,93; 10,310 Sundberg, Johan, 4,99 Swinney, David, 10,307 Thissen, David, 9,305 Thomas, Jane, 4,3 11 Tranel, Bernard, 3,35 1 Tuller, Betty, 4,303 Turnbsll, William, 8,145 Turvey, Michael T., 9,237; 10,313
343
Tversky, Amos, 7,409 Tyler, Lorraine Komisarjevsky, 8, 1 Vadhan, Vimla P., 5,251 Valian, Virginia, 4, 155; 6,1; 7,317; 10,323 Van der Molen, Hugo, 7,315 Van Duyne, P. C., 2,239 Varela, Francisco C., 1,97 Wales, Roger, 4, 155 Walker, E. C. T., 8,263 Wanner, Eric, 8,209 Wason, P. C., 3, 141 Wasow, Thomas, 4,203 Watkins, Michael J., 10,331 Weiner, Susan L., 4,3 1 Wexler, Kemsth, 6,327 Wilcox, Stephen, 3,245 Wilensky, Robert, 5,133 Wilks, Yorick, 10,337 Wilson, Deirdre, 10,281 Winograd, Terry, 5, 15 1 Wishart, Jennifer G., ! , 165 Wolf, Joan Z., 6,89 Woodruff, Guy, 7,333 Yates, Jack, 3,227 Zaidel, Eran, 8, 187 Zelazo, Philip R., 2,385 Zimbardo. Philin G.. 2.243
Cognition
Cumulatiw Contents of Voluares II to Volume 1 Editorial, 9 NOAM CHOMSKY (M.I. T., Cembridge, Mass.) Psychology and ideology. 11 THOMAS G. R. BOWER and J. G. PATERSON (University of Edinburgh) Stages in the development of the object con’cept, 47 PHILIP N. JOHNSON-LAIRD (University College, London) The three-term series problem, 57 JERRY FODOR (M.I.T.. Cambridge, Mass.) Some reflections on L. S. Vygotsky’s Thought and Language, 83 HUMBERTO R. MATURANA, FkANCiSCO G. VARELA, SAMV G. FHENK IUniversi>fy of Chile) Size constancy and the problem of perceptual spaces,97 JONAS LANGER (University of California) and SIDNEY STRAUSS (Tel Aviv University) Appearance, reality and identity, 105 LILA R. GLEITMAN, HENRY GLEiTMAN and ELIZABETH F. SHIPLEY fUniversityofPennsy/vanial The emergenceof the child as grammarian, 137 T. G. R. BOWER and JENNIFER G. WISHART lUniversity of Edinburgh) The effects of motor skill on object permanence, 165 URSULA BELLUGI and SUSAN FISCHER (The Salk lnsrjrute for P&logical Studies) A comparison of sign language and spoken language, 173 JAM ES R . LACKN ER (blassachusetis Institute of Technologyj An auditory isiurcionof depth, 201 JOHN D. BRANSFlORD (State University of New York at Stony Bnaok) and JEFFERY J. FRANKS Waederbilt
University~
The abstraction of linguistic ideas: A review, 211 DAVID PREMACK ((University of Celifomia, San& Barbara) Concordant prefferencesas a precondition for affective bL( not fir svmboiic communication (or How to do experimental anthropology), 251 DAVID LAYZER ~1yerverd Coltqw ObservatoryJ Science or superstition? (A physical scientist looks at the IQ contro&rsyj, ‘265 R. J. HERRNSTEIM (Harvard Unhwrsity) Whatever heppened to vaudeville? A reply to Professor Chomskv, 3C 1. A. N. LEONTIEV and A. R. LURIA (Moscow Univ8rsity/ Some notes concerning Dr. Fodor’s ‘Reflections on L. S. Vygotsky’s Thought and language’, 311 H. SINCLAIR fUnivars!ty of Genev8) Some comments on Fodor’s ‘Reflections on L. S. Vygotsky’s Thought 8nd 18nguage’, 317 SIDNEY STRAUSS (Tel-Aviv University) Inducing cognitive development and learning: A review of short-term training experiments. 1. The organismic developmental approach, 329 J. R. LACKNER and M. F. GARRETT &4wachusetYs lnsritufe of Technology1 Resolving ambiguiv: Effects of biasing contexa in the unattended ear, 359 J. ST, 8. T. EVANS (Sir John Cass School of Science and &hnology, City of London &ifl=hnicl On ehe problems of interpreting reasoningdata: Logical and psychologicel approac;ies.373
Cbmulotive Contents
T. SHALLICE (University College London) The Ulster depth Interrogation techniques anti their relation to sensory deprivation research, 385 NORM CHMSKY fM&sac~u&?rtsIffstitute of Technology) tXwxwnt$ on He&stein’s response,407
R. J. HERRNSTEIN fHafvard University) Comments on Professor Layzer’b ‘Science or superstition’, 419 DAVID LAYZER fH8nfard University) A rejoinder to Professor Harmstein’s comn;ents, 423 ARTHUR R. JENSEN (University of California, Berkeley) The IQ colrtroversy: A reply to Lavzer, 427 DAVID LAYZER (Hart%& iiniversity) _ Jensen’sreply: The sounds of silence, 454
Volume 2 Editorial, 7 JOHN KIMBALL (Indiana Ut?iuersity, Blo~mk?gtonl Seven principles of surface structure parsing in natural lar!guage,15 R. MILLEF (Universiisy of Witwatersrandl The use cf concrete and abstract concepts bv children and adults, 49 PHILIP BilEBERMAN (University of Connecticut, Stofrs~ On the evolution of language: A unified view, 59 C. A. PERFETTI luniversity of Pittsburgh) Ret&&al of sentence relations: Semantic vems syntactic deep structure, 95 J. P. BRONCKART and H. SINCLAIR lUniversit& de GenBvel Time, tense and aspect, 107 H. PUTNAM Marx& UnivsrrityJ Raiuctionism
&?ltdthe rutun, of PS@lolO#,
rg1
H. B. SAVIN IU~imwst%y .Tf Pennsylvania) Professorsand psychological researchers:Conflicting values in conflicting roles, 147 EVE. V. CLARK (Stanford University) Non-lingristbz strategies and the acquisition1of word meanings, 161 JEAN-PAUL BOON§ fUniver&y 0.f Paris) AcoegEebwty, intwpmt&mn ar!j knowledfp of the world: Remarks on the verb PLANTER (to plmtL 383 HARRIS 8. GAVIN {University of Pennsylvanktl Meanings and concepts: A review of Jerrold J. Katz’s Semantic theory, 213 P. C. VAN DUYNE (Uniwrsity College, London) A abrt nc)mz+ on Evans’ criticism of re&;oning experiments and his matching responsehvpothesis, PHILIP G. ZlBlBARDO /Stanford University/ On the ethii of intervention in human psychological research: With special reference to the Stanford prison experiment, 2413 HARRIS B. SAVIN lUniversity of PbnnsylvanisJ Ethics for godsand men, 267 DAVID CAPLAN iMcGil/ Medical &h&J A wte on the abswact read& of verbs of perception. 269 LOUIS M. GOLDSTEIN and JAMES R. LACK.NER fhndeis Univefsity~ Alterations of the phonetic coding of speech sounds during repetition, 279
Cognition
JONATHAN BAF’ON lMctl#aster University) Semantic components and conceptual development,
347
299
KENNETH I. FORSTER and ILMAR OLBREI fMonash University) Semantic heuristics and syntactic analysis, 319 CHARLES J. BRAINERD /University of Alberta) Neo-Piagetian training experiments revisited: stage hypothesis?, 349
Is there any support for the cognitive-development
A. SCHAERLAEKENS (University of Leuvenl A generative transformational model for child language acquisition, 371 PHILIP R. ZELAZO, J. ROY HOPKINS, SANDRA JACOB”ON and JEROME KAGAN Univ ?rsityj Psychological reactivity to discrepant events: Support for the curvilinear hypothesis, 385
,lHarvard
DEREK EDWARDS 0Jniversity of Sussex) Sensoory-motor intelligence and semantic relations in early child grammar, 395 KEITH E. NELSON and JOHN D. BONVILLIAN (Stanford University) Concepts and words in the lB-monthold: Acquiring concept names under controlled cor7ditions. 435 0. TERENCE LANGENDOEN (City University of New York), NANCY KALISH-LANDON bLoyo/a University) and JOHN DORE fCity University of New York) Dative questions: A study in the relation of acceptability ‘:_:grammaticality of an Er;:lish sentence type, 451 STEVEN P. R. ROSE (The Opm University) and HILARY ROSE (London School of Economics) ‘Do not adjust your mind, there is a fault in realitv’ - Ideology in neurobiology, 479
Volume 3 JONAS C-ANGER (University of California, Berkelyj lneeractional aspects of cognitive organization, 9 PIERRE MOUNOUD (University of Geneva) and T. G. R. BOWER (University of Edinburgh) Conservation of weight in infants, 29 JENN: FER HUGHFS (Medical Research Council Developmental Psychology lJn,‘t, Lo&on) Acquisition of a ,lon-vocal ‘language’ by aphasic children, 41 TENON W. PYLYSMYN (University of Wastern Ontario) Minds, machines and phenomenology : Some reflections on Dreyfus’ What computers can’t do, 57 ROGER SCH NAITTC R (Illinois Wesleyan University) The decline of reason, 79 THOMAS G. BEVER iColumbia Reason and un-reason, 83
University) and JACQUES
URSULA BELLIJGI (The Se/k Instituttv~. EDWARD and PATRICIA SIPLE (University of Rochester) Remembering in signs, 93
MEHLER
S. KLIMA
(C.N.R.S.)
/University
of CMfurnia,
JOSE MORAIS (Universiti libre de Bruxellesj The effects of ventriloquism on the right-side advantage for verbal material, 127 P. C. WASON (University College London) and J. ST. B. T. EVANS Dual processes in reasoning?, 141 SIDNEY STRAUSS ITel-Aviv A reply to Brainerd, 155
0Vymouth
Polytechnic)
University)
ANNETTE KARMILOFFSMITH and BARBEL INHELDER “If you want to get ahead, get a theory”, 195
(Univvrsity of
Geneva!
San Dies$
Cumuktive Contents
348
DANIEL N. OSHERSON (University of Pennsylvania) and ELLEN MARKMAN Language and the ability to evaluate contradictions end tautologies, 213 CATHERiNE GARVEY, ALFONSO CARAMAZZA and JACK YATES SitYl Factors influencing assignment of pronoun antecedents, 227
(University of Illinois)
(The Johns Hopkins Univer-
STEPHEN WILCOX and DAVID S. PALERMC‘ (The Pennsylvania State University) ‘in’, “on’, and ‘under’ revisited, 245 J. S. BR!UNER (University of Oxford) From communication to language - A psvchofogical perspective, 255 ANDREii LUGG Wniversity of Ottawa) Putnam on reductionism, 289 HI LAR Y PUTNAM Harvard Reply to Lugg, 295
University)
ULRiC NEISSER and JOHN A. HUPCEY fCorne// University) A Sherlockian experiment, 307 ARNOLD L. GLASS and KEITH J. HOLYOAK (Stanford University) Alternative conceptions of semantic theory, 313 JAMES McGARRIPLE end MARGARET Conservation aL*-lents, 341
DONALDSON
(University of Edinburgh)
SANFORD A. PQiANE, BERNARD TRANEL and HARLAN LANE lllnlversity San Diego) On the psychological reality of a natural rule of syllable structure, 351
of CafifornB
at
PHILLIP SHAVER, LEE PIERSON, and STEPHEN LANG (Columbia University, New York City) Converging evidence for the functional significance of imagery in problem solving, 359 A. R. LURIA Wniversity of &scowl Scientific perspectives and philosophical dead elnds in modern linguistics, 377 J. St. 6. T. EVANS (Plymouth Polytechnic) On interpreting reasoning data - A reply to Van Duyne, 387
Volume 4 Editorial, 7 KATHERINE NELSON /Yale UniversiQ.1 Some attributes of adjectives usad by young children, 13 SUSAN L. WEINER (Educationa/ Testing Swvi&?l and HOWARD New York) Ocular motility and cognitive process, 31 EWVARD S. KLIMA flJnivtWty of Wtiwnia et Snn D&O) Institute for Bii?logica Studies) Poetry and song in a language without sound, 45 JONAN
SUNDBERG
IRoya/
lnniture of Ttidobgy,
EHRLICHMAN
and URSULA
Stockholml and BJ6RN
(City University af
BELLUGI
LINDBLOM
(The Wk
(Stock.
h@h Universityj _ Generative theories in language and music description, 99 ROGER BROWN ftfarvard Unive&yl Referenoe - In memorial tribute to Eric Lenneberg, 125 VIRGINIA VALIAN (CUNY Gmd~te Cantar, lllrw Ywk) and ROGER WALES llJnbe&y of St Andnmr) What’s !&ht: talkers help listeners hpar and understand by clarifying sentential relations, 155
Cognitiffn 349
MICHAEL T. MOTLEY and BERNARD J. BAARS /University of California) Semantic bias effects orbthe outcomes of verbal slips, 177 SUSAN GOLDIN-MEADOW,
MARTIN
E. P. SELIGMAN
and ROCHEL GELMAN
(University
of
Pennsylvania)
Language in the two-year old, 189 DANIEL N. OSHERSON ilJniversity of Pennsylvania) and THOMAS WASOW f%anford Universityj Task-specificity and species-specificity in the study of language: A ~methodologicalnote, 203 ELIZABETH SPELKE, WILLIAM Skills of divided attention, 215
HIRST, and UI..RIC NEISSER ICornell University, New York)
HUGH FAIRWEA iER (Universities of Oxford.and Bologna) Sex differences in cognition, 231 THOMAS H. CARR (George Peabody College, Nashvillej and VERNE
R. BACHARACH
(Acadia
University. Wolfvillej
Perceptual tuning and conscious attention: Svstems of input regulation in visual information processing, 281 JAMES R. LACKNER Mrandeis University and Massachusetts lnstitutc of Technologyj, and BETTY TULLER (Brandeis University, The influence of syntactic segmentation on perceived stress,303 JOHN MORTON (MRC Applied Psychology Unit, Cambridge) On recursive reference, 399 MATTHEW HUGH ERDELYI andSHlRA FINKELSTEIN IBrooklyn Co/lose), NADEANNE HERRELL, BRUCE MILLER and JANE THOMAS (The Stare University, Rudgersj Coding modality vs. input modality .in hypermnesia: Is a rose a rose a rose?, 311 B. ELAN DRESHER (University of Mas~chusettsj, NORBERT HORNSTEIN (Harvard University) On some supposed contributions of artificial intelligence to the scier 1ific study of language, 321
Volume 5 MARC JEANNEROD IIMSERM Lyon1 Sur la nature de la connaissanceempirique, 3 A tribute to H. L. Teuber (1916-1977) MARY-LOUISE KEAN &fassachussetrsInsriruta of Technologyj The linguistic interpretation of aphasicsyndromes: Aggrammatism in Broca’s aphasia,an example, 9 GIYOO HATANO (Dokkyo University, &pan), YOSHfO MIYAKE (Nati~naf tnstitute for Educational Research, Japan) and MARTIN G. BINKS (The university of Liverpooij Performance of axpert abacus operators, 47 SANDY PETREY (State University of New York lbStony Brook) Word associationsand the development of lex Sal memory, 57 ROGER BROWN end JAMES KULIK 0farva d UnhwsWl Flashbulb memories, 73 FRANCOIS GROSJEAN and HARLAN LANE fNohheastern Universityj Pausesand syntax in American sign language, 101 DEAN DELIS and ANNE SAXON SLATER (University of Wyomingj Toward a functional theory of reduction transfprmations, 119 ROGER C. SCHANK and ROBERT WfLENSKY (Yale University) Response to Dresher and Homstein, 133 B. ELAN DRESHER (Unfwrsfry of M~husatts, Amherst) and NORBERT HORNSTEIN (Harvard Univarsityj
Reply to Schank and Wilenrkv. 147
Cumulative Contents
350
TERRY WINOGRAD Branford University) On some conteSted suppositionsof generative linguistics about scientific study of language, 151 PETER SCHdNBACH ff?uhr-Universitir Bochum) In defense of Roger-Brown einst himself, 184 ROGER BROWN (Harvard University) In reply to Peter SchQnbach, 185 PHILIP N. JOHNSON-LAORD fliniversify of Sussex) Procedural semantics, 189 BV. PATRICK DICKSON si’ry of Tokyo) Referential relativiw:
lU;nivers;ty of Wisconsinl NAOMI
MIYAKE
and TAKASHI
MUTO fUniver-
Culture-boundedness of analytic and metaphoric communications,
215
ROBERT GRIEVE, ROBERT HOOGENRAAD and DlARMiD MURRAY &/niversify of St. Andre& On the young child’s use of Iexis and syntax in understanding locative instructions, 235 VIMLA P. VADHAN and DANIEL Attention and cognition, 251
W. SMOTHERGILL
&yracuse University)
J. St. B. T. EVANS and S. E. NEWSTEAD iPlymouth PolyfecE?nic) Language and reasoning: A study of temporal factors, 265 RICHARD BYRNE (Universiry of St. Andraws Planning meals: Problem-solving on a real data-base, 287 ARTHUR S. REBER and SELMA LEWIS (Brooklyn College of CUNY) Imp!!& kerning: An analysis of the form and structure of a body of tacit knowledge, 333 E. J. 7IOBlNSON Development
and W. P. ROBINSON (Macquarie University, Australia) in the understanding of causes of success and failure in verbal communication,
B. ELAN DI:ESHER l&own Reply to Winogred, 377
JlJJlJ;U$NZEMER Transformations,
Universiryj and NORBERT
MAYER,
ANNE
ERREICH
HORNSTEIN
and VIRGINIA
363
(Harvard University)
VALIAN
(CUNY Graduate Center,
basic operations and language acquisition, 1
W. A. PHELLIPS, S. B. HOBBS and F. R. PRATT &Wing
University, Stirling1
Intellectual realism in children’s drawings of cubes, 15 MARYLIN FORD and VIRGINIA M_.HOLMES (University of Melbourne, Parkville. Vie.; Planning units and syntax in sentence production, 35 WILMA BUCCI (State UnFfersity of New York, Downstate Medical Center) The interpretatior! ;,i universal sffirmative propositions, 55 P. J. HAMPSON and P. E. MORRIS lUniversify of Lancaster, lancasrer1 Unfulfilled expectations: A criticism of Neisser’s theory of imagery, 79 MARC H. BORNSTEIN, CHARLES G. GROSS and JOAN 2. WOLF (Princeton University) Perceptual similarity of mirror images in infancy _89 ELLEN H. GROBER, WILLIAM BEARDSLEY an j ALFONSO University) Parallel function strm in pronoun assignmenr. 117
JOHN M. SORENSEN.
WILLIAM E. COOPER ar4d JEANNE Tmhnoiogyj speech timing of grammatical categories, 135
CARAMAZZA
M. PACCIA
(The Johns Hopkins
Wssachuserrs
WERNER DEUTSCH ltWx-P/anckCesWzhaft~ and THOMAS PECHMANN lahnl Ihr, dir, or mirf On the acquisition of pronouns in German children, 155
(Universitit
lnsfitufe of
Marburg/
Cognition 35 1
ULRIC NEISSER (Cornell University) Anticipations, images, and introspection,
169
JOHN L. LOCKE i/ns?itufe for Child tj)ehevior end Development, Champeign, 111.1 Phonemic effects in the silent reading of hearing and deaf children, 175 ARTHUR S. REBER (Brooklyn College of CUNY) ,md RHIANON ALLEN (ClJ/VV Graduete Centre) Analogic and abstraction strategies in synthetic grammar learning: A funct,ionalist interpretation, l&9 DAVID NAVON (University of Heife) On a conceptual hierarchy of time, space, and other dimensions, 223 J. A. FODOR (M8ss8chusetts Institute of Technology) Tom Swift and his procedural grandmother; 229 P. N. JOHNSON-LAIRD (Laboratory nf Experimental Psychology, University of Sussex) What’s wrong with Grandma’s guide to Procedural semantics: A reply to Jerry Fodor, 249 DANIEL N. OSHERSON (~8ss8chusetts institute of Technologyl Three conditions on conceptual naturalness, 263 LYN FRAZIER and JANET DE.AN FODOR (University of Connecricurl The sausage machine: A new two-stage parsing model, 291 KENNETH WEX LER lUniversiry of Celifornia, lrvinel A review of John R. Anderson’s Language, Memory, and Thought, 327 HERMAN H. J. KDLK (University off [imegen, The Netherlands) The linguistic interpretation of Broca’s aphasia. A reply to M.-L. Kean, 353
VDCm?@a Editorial,
1
MARY SUE AMMON and DAN I. SLOBIN (University of Californie, Berkeley) A cross-linguistic study of the processing of causative sentences, 3 ANTHONY F. JORM (Deekin University, Austr8lie) The cognitive and neurological basis of developmental review, 19
dyslexia:
A theoretical
framework
a&
HUGO VAN DER MOLEN and JOHN MORTON lMRCApp/ied Psychology Unit, Cambridge) Remembering plurals: Unit of coding and form of coding during serial recall, 35 ANNE CUTLER and JERRY A. FODOR (~8ss8chusett /nstitu.Ze of Technology1 Semantic focus and sentence comprehension, 49 JOHN KLOSEK (Gredwete Center, CUNY) Two unargued linguistic assumptions in Kean’s “phonnlogical”
interpretation
of agrammatism, 61
MARY-LOUISE KEAN (University of Celifornie, Irvine) Agrammatsim: A phonological deficit?, 69 HELEN GOODLUCK (University of Wisconsin, Medisonj M8ss8ChUSetts, Amherst) A reevaluation of the basic operations hypothesis, 85
and LAWRENCE
SOLAN
JERRY A. FODOR (M8S8chusetts Institute of Technology) In reply to Philip Johnson-Laird, 93 THOMAS R. SHULTZ, ARLENE DOVER and ERIC AMSEL lMcGi// University) The logical and empirical bases of conservation judgements, 99 ANAT NINIO (The Hebrew University, Jenrselem) Piaget’s theory of space perception in infancy, 125
(University
cf
Cumulative Contents
FRANCESCO ANTINUCCI (CNR, Rome), ALLESSANDRO DURANTI (University of Rome) and LUCYNA GEBERT lUniversity of Genoa) Relative clause strqcture, relative clause perception, and the change from SOV to SVO, 145 MARK S. SEIDENBERG (Columbia Univafsity) and LAURA A. PETITTO (New York UniVafsitY) Signing behavior in apes: A critical review, 177 STEVEN PINKER (Harvard University) Formal models of language learning, 217 TIMOTHY E. MOORE (Glendon Co/&e, York University) and IRVING BIEDERMAN lState University of New York) Speeded recognition of ungrammaticality : Double violations, 285 STEPHEN P. SCHWARTZ flthaca Col/ege) Natural kind terms, 301 ANNE ERREICH, JUDITH WINZEMER MAYER and VIRGINIA VALIAN (CUNY Graduate Centef) Languageacquisition hypotheses: A reply to Goodluck and Solan, 317 JOSE MORAIS, LUZCARY,JESUSALEGRIAandPAULBERTELSON iUniversitiLibredeBruxe/kw~ Does awarenessof speech as a sequenceof phones arise spontaneously?, 323 GUY WOODRUFF and DAVID PREMACK Wniversity of Pennsy/vanial Intentional communication in the chinlpanzee: The development of deception, 333 ,‘, LANGFORD and V. M. HOLMES (University of Melbourne) Syntactic presupposition in sentence -.amprehension, 363 L. JONATHAN COHEN (Oxford Univer&:yj On the psychology of prediction: Whose is the fallacy?, 385 DANIEL KAHNEMAN Wniversity of British CoJumbiaj and AMOS TVERSKY &anfofd Universityl On the interpretation of intuitive probability : A reply to Jonathan Cohen, 499 ANDREW W. ELLIS (Univwsity of Lancaster) Develapmental and acquired dyslexia: Some observations on Jorm (19791,413 ANTHONY F. JORM (Damkin Universityj The nature of the reading deficit in developmental dyslexia: A reply to Ellis, 421
Volume 8 WILLIAM MARSLENWILSON and LORRAINE KOMISARJEVSKY TYLER (Max-P&nck-fnstitut f6r Psycholinguistik, N&genj The temporal structure of spoken languageunderstanding, 1 JOHN R. ANDERSON fCarnt@e-Me//on University) On the merits of ACT and information-processing psychology: A response to Wexler’s reviaw, 73 L. JONATHAN COHEN (Oxford University) Whose is the fallacy? A rejoinder to Daniel Kahneman and Amos Tversky, 89 MICHAELSTUDDERT-KENNEDY (QueensCollegeandGreduataCentar, CUNM Languageby hand and by eve. A review of Edward S. Klima and Urt!ula Bellugi’s, The Signs of Language,93 HERBERT H. CLARK and DALE H. SCHUNK &anford University) Polite rasponfasto polite raquasts, ii 11 LANCE J. RIPS (Univers?ty of Chicago,land WILLIAM TURNBULL l&non Fraser University) How big is big7 Relative and absolute properties in memory, 145 RHIANDN ALLEN (Tha Gmduata Center of CUNY) and ARTHUR S. REBER 43fooklynCollege of CUNY) Very long term manory for tacit knowledge, 175
Cognition
ANN M. PETERS (University of Hawaii) and ERAN ZAIDEL and Division of Biology, California Institute of Technology1
(University
353
of California, Los Angeles
The acquisition of homonymy, 187 ERIC WANNER (Sussex University) The ATN and the SausageMach ne: Which one is baloney?, 209 ELLEN M. MARKMAN,
MARJORIE
S. HORTON and ALEXANDER
G. McLANAHAFJ (Stanford
University)
Classes and collections: Principles of organization in the learning of hierarchical relations, 227 N. H. FREEMAN, S. LLOYD end C. G. SINHA (Department of Psychology, University of Bristol) Infant search tasks reveal early concepts of containment and canoi>ical usage of objects, 243 J. A. FODOR, M. F. GARRETT, E. C. T. WALKER and C. l-l. PAFIKES (Psycho!ogy Department, Mawhusetts
Institute of Technology)
Against definitions, 263 JOHN H. FLAVELL, SUSAN 3. SHIPSTEAD and KAREN CROFT (Stanford University) What young children think you see when their eyes are closed, 369 HENRY HAMBURGER INationalScience Foundation) A deletion ahead of its time, 389 JANET DEAN FODOR (Universityof Connecticut) and LYN FRAZIER (University of Massachusetts) Is the human sentence parsinjl mechanism an ATN?, 4’17
Volume 9 ULRIC NEISSER (Cornell University) John Dean’s memory : A case study, 1 MATTHEW HUGH ERDELYI and JUDY B. STEIN (Brooklyn College of CUNY) Recognition hypermnesia: The growth of recognition memory (d’l over time with repeated testing, 23 DANIEL N. OSHERSON Massachusetts Institute of Technology) and EDWARD E. SMITH (Stanford University and Bolt Beranek and Newman, Inc.) On the adequacy of prototype theory as a !neory of concepts, 35 GILLIAN COHEN (University of Oxford) Inferential reasoning in old age, 59 THOMAS H. CARR (Michigan State Univarsity) Building theories of reading ability: 3n the relation between individual differences in cognitive skills and reading comprehension, 73 ALFONSO CARAMAZZA, MICHAEL McCLOSKEY and BERT GREEN (The Johns Hopkins University) Nal’ve beliefs in “sophisticated” subjects: misconceptions about &ectories
N. R. IBBOTSON /Lambridge
Medical School)
and JOHN MORTCN
of objects, 117
ftl4.R.C. Applied Psychology
Unit, CambrIdge)
Rhythm dnd dominance, 125 J. A. FODDR l#asmchusetts Institute of Technology)
and Z. W. PY LYSHYN iUniversity of Weste, -
Onteriol
How direct is visual perception?: Some reflections on Gibson’s “Ecological Approach”, 139 KENJI HAKUTA fYa/e University) Grammatical description versus configurational arrangement in language aquisition: The cese of relative clausesin Japanese, 197 M. T. TURVEY &Jniversity of cbnnecticut, Sfvrrs, C4nnecticutand HaskinsLaboratories, Nev;Heven, Connecticut), R. E. SHAW (University of Connecticut, Storrs, Connecticutl, E. S. REED @eerier
CIrmulutive Contents
354
for ReseamtiinHuman Learning, Minneepolis. Minnesvta~ and W. M. MACE (Trinity College, Hartford, Connecticut1 Ecological laws uf perceiving and acting: In reply to Fodor and Pylyshyn (19811,237 SUSAN KEEPER and DAVID THISSEN (University of Kansas) How polite?: A reply to Clark and Schunk, 395 HERBERT l-f. CLARK (Stanford University) and DALE H. SCHUNK Politeness in requests: A rejoinder to Kemper and Thissen, 311
(University of Houston)
Volume 10 Editorial, 1 TONY ADES &?rigMon. Englandl Time for a purge, 7 ALAN BADDELEY (MRC Applied Psychology Unit, Cambridge) The concept of ‘working memory: A view of its current state and probable future development, 17 SHEILA E. SLUMSTEiN i&own University) and KENNETH of T’hnc 3gy) Phcnatrc features and acoustic invariance in speech, 25
N. STEVENS
lMas.sacItusetts InstitMe
ALBERT S. BREGMAN (McGi// University) Chomrky without language, 33 JOAN BRESNAN ikkssachusetts Institute of Technology) An approach to universal grammar and the m#ental representation of language, 39 DONALD E. BROADBENT lUniversity of Oxford) Seieztive and control processes, 53 DAVID CAPLAN &%tawa Civic Hospital) Prospects for neurolinguistic theory, 59 ANNE CUTLER /University of Sussex) Making up materials is a confounded experiments at al in 1999?, 95
nuisance, or: Will we be able to run any psycholinguistic
JAMES E. CUTTING fComt#/ Universitv) Six tenets for event perception, 71 PETER 0. EIMAS (Brown University) Infants, speech, and language: A look at some connections, 79 GILLES FAUCGNMIER Wniversit~de Paris/ Pragmatic function; and mental spaces, 85 IRA FISCHLER Wniversity of Florida) Research of context effects in word recognition: Ten years back and forth, 89 MERRILL F. GARRETT Wssachusetts Institute of Technology) Objects of psycholinguistic enquiry, 97 LILA R. GLEiTMAN lllniversity ofPennsylvania Maturational determinants of language growth, 103 0. W. HAMLYN @irkbeck Q//age, University of London1 Cognitive systems, ‘folk psychology’, and knowledge, 115 ALICE F. HEALY IYak Unirerrty) cogcfiitkre processes in reading text, 119 JULfAN HOCHBERG ICo/umbia University) Or3 cognition in perception: Perceptual coupling and unconscious inference, 127 MARC JEANNEROD fl,/V.S,E.R.M., Lyon) Specialized channels for cognitive responses, 135
Ciqfnition 355
PHILIP N. JOHNSON-LAIRD fUniversity of Sussex) Cognition, computers, and mental models, 139 JOHN JONIDES and DAVID Capturing attention, 145
E. IRWIN
lUniversity of Michigan)
ANNETTE KARMILOFF-SMITH (Sussex University and Max-Plan&-lnstitut Getting developmental differences or studying child cevelopment?, 151
fiir Psycholinguistik)
FRANK KEI L (Cornell University) Children’s thinking: What never develops?, I59 WILLIAM KESSEN (Yale University/ Early settlements in New Cognition,
I67
STEPHEN MICHAEL KOSSLYN fBrandeis University) Research on mental imagery: Some goals and directions, I73 JONAS LANGER (University of California, Berkeley) Logic in infancy, 181 WI LLEM J. M. LEVELT Dg& vu?, 187
(Max-PAanck-lnstitut
for Psycholinguistikl
ELIZABETH F. LOFTUS Wniversity of Washingron) Natural and unnatural cognition, 1!33 H. CHRISTOPHER LONGUET-HIGGINS (Universityofsussex) Artificial intelligence - a new theoretical psychology?, 197 DAVID McNElLL (Universiry of Chicago) Action, thought and language, 201 JOHN C. MARSHALL and FREDA NEWCOMBE {The Radcliffe lnfirmary, Lexical access: A perspective from pathology, 209
Oxford)
GEORGE A. MILLER {Princeton University) Trends and debates in cognitive psychology, 215 JOHN MORTON lMf?C Applied Psychology Unit, Cambridge) Will cognition survive?, 227 DONALD A. NORMAN and DAVlD E. RUMELHART (University of California, San Diego) The LNR approach to human information processing, 235 DANIEL N. OSHERSON hllassachusetts institute of Technology) Modularity as an issue for cognitive science, 241 STEVEN PINKER (Harvard University) What spatial representation and language acquisition don’t have in common, 243 DAVID 8. PI SON I (Indiana University) Some current theoretical issues in speech perception, 249 MICHAEL I. POSidcR (University of Oregon) Cognition and neural systems, 261 ZENON W. PY LYSHYN Wniversiry of Western Ontariol Psychological explanations and knowledge-dependent
processes, 267
DAN I, S LO8 I N (University of California, Berkeley) Psychology without linguistics = language without grammar, 275 DAN SPEREER 1C.N. R.S. and Universith London) Pragmatic% 281
de Paris) and DEIRDRE
JOHN E. R. STADDON (Duke Universifyl Cognition in animals: Learning as program assembly, 287 SIDNEY STRAUSS (Ttrl Aviv University) Cognitive development in school and out, 295
WILSON
(UniversitY
Cole,
Cumuhtive
356
Con
tents
MICHAEL STUDDERT-KENNEDY (Gueens College and Graduate Center, CUNY and Haskins Labontories! The emergence of phonetic structure, 301 DAVID SWINNEY (Tufts University) The prmsof language comprehension; an approach to examining issues in cognition and language, 307 MICHAEL T. TURVEY (University of Connecticutandlfaskins Nhhersity of Connecticud Cognition: The view from ecological realism, 313
Laboratories) and CLAUDIA
VIRGINIA VALIAN (Columbia University) Linguistic knowledge and language acquisition, 323 MICHAEL J. WATKINS (Rice University) H!~man memory and th information-processing
metaphor, 331
YORICK WILKS (Univershy of Essexj A position note on natural language understanding and artificial intelligence, 337
CARE LLO