Cognition, Vol. 9, No. 1

Editorial Board Hiroshi Azuma Faculty of Education, Tokyo University, Hongo, Bunkyo-ku Tokyo, Japan Paul Bertelson Laboratoirede &ychologie Exp&imentale, UniversitiLibre de Bruxelles 117 Av. Adolphe Buyi, B-I 050 Brussels,Belgium Manfred Bierwisch Akademie der Wissenschaften der DDR, Zentralinstitutfir Sprachwissenschaft, Otto Nuschke Strasse22123 108 Berlin, D.D.R.

Eve Clark Department of Linguistics, Stanford University, Stanford, CalifI94305, U.S.A.

Jerry Fodor Dept. of Psychology, M.I.T. El O-034, Cambrrdge,Mass.02139, U.S.A.

Anne Cutler Laboratoryof Experimental Psychology, Centre for Research on Perception and Cognition, Universityof Sussex, Brighton BNl, Ct. Britain

Kenneth Forster Dept. of Psychology, Monash University, Clayton, Vie. 3168, Australia

James E. Cutting PsychoIogyDepartment, UrisHall, Cornell University, Ithaca, N. Y. 14853, USA.

Peter D. Ennas WalterS. Hunter Laboratory Ned Block of Psychology. Dept. of Philosophy,M.I.T., Cambridge,Mass.02139, U.S.A. Brown University, Providence,R.I. 02912, U.S.A. Melissa Bowerman PsychologyDepartment, Gunnar Fant Universityof Kansas, Lab. of Speech ~ansmission. Lawrence, Royal Instituteof Technology, Kansas66044, S-I 0044 Stockholm 70, Swedpn &%A. Francois Bresson Laboratoirede Psychologie, 54 bvd. Raspail, F- 75006 Paris,France Roger Brown Dept. of Psychology, HarvardUniversity, ambridge* Mass.02138, Lr.S.A.

Giues Fauconnier 9 Rue des Guillemites, 75004 Paris,France David Fay Bell Laboratories, Warrenville- NapervilleRoad, Naperville, 111*ao540* U.s-A.

David Caplan Divisionof Neurology. QttawaCtvicHospital, Ottawa,Ont. KlS 2A3, Canada

Ira Fischler Department of Psychology, Universityof Florhia, Gainesville,Fla. 3261 I, U.S.A.

Noam Chomsky Dept. Modern Languagesand Linguistics,M.I.T., ambridge, Mass.02139, U.S.A.

Janet Dean Fodor DepartmentofLinguistics, .Universityof Connecticut, Storrs, corm. 06268, CIS.A.

MerrdlGarrett Department of Psychology, M.I.T. El O-034, Cambridge,Mass. 02139, U.S.A.

Lila Gbitman GraduateSchool of Education, University of Pennsylvania, 3700 WalnutStreet. Phdadelphia,Pa. 19104. U.S.A. David T. Hakes, Department of Psychology, Umversltyof Texas, Austm, Td-7.78712, USA. Henry Hecaen Directeur d ‘Etudes, Ecole hatigue des HautesEtudes, Unit6 de Recherchts Neuropsycholo@ues, I.N.S.E.R.M.. 2, rued’Al&a, F- 75014, Paris,France Mlchel Imbert Laboratoirede Neurophysiologic, Coll$gede France, 1I PlaceMarcelinBerthellrot, F- 75005 Paris,France B&be1 Inhelder Faculti de Psychologieet des Sciences de I’Educatiorr, Universitkde GenpVe. CH-1211 Geneva14, Swirzerland Marc Jeannerod Laboratoirede Neuropsychologi! Exp&imeneale, 16 Av. Doyen L&phe, F-69500 i&on, France

Philip Johnson-Laird Laboratoryof Experimental Psychology, Centre for Researchon Per ception and Cognition, Sussex Uni9ersity. Bright I 2Nl PQG, Ct. Britain Peter W. Jusczyk Dept. of Psychology, Vniwersityof Oregon, Eugene, Oreg. 97403, VS.A. Jenold J . Katz Dept. of i&g&tics, CUNY GraduateCenter, 23 W42nd Street, New York, N. Y. 10036. U.S.A. Mary-LouiseKean CognitiveScience fiogram, School of SocialSciences. Universityof iMifomia, Irvine, Gali) 92717, U.S.A. Frank Keil PsychologyDepartment, CorneB Vni~ersity, Ithaca*N. Y. 14853, U.S.A. Edward Klb Ia Dept. of Linguistics,La Jolla, University0 f Califom~, &m Diego, Calif 92037, U.S.A. Stephen M, %.osslyn Department of Psychologyand SocialRrhztions, HarvatdVisi9ersity. Wil&m Jones Ihdl, 33 KirH nd Street, Cqmbkge. Mass.02138, USA.

John Lyons Llept. of Linguistics, AidamFergusGnBuilding, Edinburgh EH8 PLL, Gt. Btitain

David Premack PsychologyDepartment, Universityof Pennsyl9anC 3813-15 WalnutStreet, Philpdelphti,Pa. 19174, &%A.

David McNeill Repartment of Behavioral Sciences, Committeeon Cognitionand Communication. Vniversityof Chicago, 5848 South UniversityAvenue, Chicago,Ill. 60637, U.S.A.

&non Pylyshyn Dept. of Psychology, The Vniversttyof Western Ontario, London 72, Ont., &mada

Wilham Marslen-Wilson l Max PlanckInstitutfir Psycholinguistik, Berg en Dalseweg79. Nvmegen, The Netherlands John Marshall NeuropsychologyUnit, Radcliffe Inftrmary, WoodstockRoad, Oxford OX2 6HE, Gt. Britain Josi Mottus Laboratoirede Psychologie Expt%imentale, VniversiteLibre de Btuxelles, I I 7 Avenue Adolphe Buyl, E-IO.50Brussels,Belgium John Morton MRCApplied Psychology Unit, I5 ChaucerRoad, CambridgeCB2 2EF, Gt. Brttain Noizet kil~oratoirede Psycholog e, 28 rue Serpente, 75006 Paris,France George

Harlan Lane Departmentof Piychology, Northevstem University, 360 Hun;kgton Avenue, Boston, Mas&&‘2115.U.S.A.

Daniel Osherson 2OC-124 (DSRE), M.I.T.. Cambridge,Mass02 l39, V S.A.

W#&em Levelt E&x ?%nck Institutfir Ayelaolhlguistik, N@negeh, The Netheriimds

Michael Posner Dept. of Aychology, Vni9ersityof Oregon* Eug,ene, Ore. 9 7403, U.S.A.

Andre Roth Lecours HGtel-Dieude Montreal, 3840 rue St. Urban, Montreal, Quebec H2W 1 T8, Canada Steven Rose Biology Department, The Onen Vniversitv. __ WaltonHall, MiltonKeynes MK 7 6AA, Ct. Britatn Scania de Schiinen Laboratoirede Psychologie, 54 BoulevardRaspail, 75270 ParisCkdex 06, France Tm Shallice MRC Applied Psychology Unit, 1.5 Chaucer Road, CambridgeCB2 2EF, Gt. Britain Dan I. Slubin Department ofPsychology, __ University of Califomta, Berkeley, Calif: 94720, V.SA. Elizð Spelke PsychologyDepartment, Universityof Pennsylvania, 3815 WalnutStreet, Phikdelphia, Pa 19104, U.S.A. MarkSteedman Department of Psychology, Un&twityof Won&k, Couentry CY4 7AL, Gt. B&in Sidney Strauss Department of Ifdueattonal Mences, Tei Aviv Vniverstty, RamatAvid, Ismel

Michael Studdert-Kennedy Department of t%mmunication Arts and Sciences, Queens collrge. CXy Universityof New York, Flushing, N. X 11367, U.&X

Edward Walker MI. Z Centerfor Cognitive Science, 77 MassachusettsAvenue, Cambridge,Mass.02139, USA.

Delrdre Wilson Department of Phonetics & Linguistics, UniversityCollege London, Cower Street, London WCIE 6BT, Ct. Britain

David Swinney Department of Psychology, 7ufIs University, Medford, Mass.02155, USA.

Peter Waoon ~ycholfnguistics, UniversityCollegeLondon, ReseaMh Unit, 4 Stephenson Way, London NW1 2HE, Ct. Britain

Edgar Zurif AphasiaResearch Center, Boston UniversityMedical Center, 150 South Huntington Avenue, Room Cl55, Boston, Mass 02130, U.S.A.

Virginia V&an 221 I Bmdway, New York, N. Y. 10024, U.S.A.

Ken Vl’exler School of SocialSciences, Universityof Gzlfornia, Irvine, Ci~lif92717, U.S.A.

Hemdna Smclab de Zwart Centte d%pistt!moIogie G&tique, Universitede GenPve, CB-1211 Geneva,SwitzerIand

Cognition,9 (1981) l-22 @ Elsevier Sequoia !%A.,Lausanne - Printed in the Netherlands

John Dean’sMemory: A casestudy ULRIC NEISSER” Cornell University

Abstract John Dean, the former counsel to President Richard Nixon, testified to the Senate Watergate Investigating Committee about conversations that later turned out to have been tape recorded. Comparison of his testimony with the actual transcripts shows systematic distortion at one level of analysis combined with basic accuracy at another. Many of the distortions ref7ected Dean’s own se@image; he fended to recall his r(ole as more ccn:ral than it really was. Moreover, his memory for even the ‘gist” of conversations was quite poor except where that gist had been rehearsed in advance or frequently repeated. But while his testiinony was often wrong in terms of the particular conversations he tried to describe, Dean was.fundamentally right about what had been happening: the existence of a ‘kover-up” and the participation of various individuals in it. His testimony was accurate at a level that is neither ‘kernan tic ” (since he was ostensibly describing particular episodes) nor %pisodic” (since his accounts of the episodes were often wrong). The term “kepisodic” is coined here to describe such memories: what seems to be a remembered e,=risodeactually represents a repeated series of events, and thus reflects a genuinely existing state of affairs.

“Have you always had a facility for recalling the details of conversations which took place many months ago ?.” Senator Inouye of Hawaii asked this question of John Dean with more than a trace of disbelief. Dean, the former counsel to President Richard M. Nixon, was testifying before the “Watergate” Committee of the United States Senate in June, 1973. His testimony had opened with a 245page statement, in which he described literally dozens of meetings that he had attended over a period of several years. The meetings were with John Mitchell, Robert Haldeman, Charles Colson, Gordon Liddy, and others whose names became American household words as the Watergate scandal brought down the Nixon administration. Some were with Nixon himself. Dean’s testimony seemed to conf!!n what many already suspected: *‘IIre first annual Dclos D. Wickens Lecture, presented at Ohio State University on October 14, 1980. Requests for reprints should be sent to Ulric Neisser, Psychology Department, Uris Hall, Cornell University, Ithaca, NY 14853 &A.

.2

UlricNeisser

that these high officials were engaged in a “cover-up*’ of White House involvement i.u the o&ind Watergate burglary. But was he telling the truth? HOW much did he re3uy remember? In a psychological experiment, it is relatively easy to determine whether what the subject says is true. The experimenter knows what really happened because sh.e staged it in the first place, or because she kept a record with which the subject’s report can be compared. Because life does not keep such records, legal testimony is usually evaluated in more indirect ways: corroborative witriesses, crossexamination, circumstantial evidence. For some of Dean’s testimony, however, it is now possible to compare what he said with a factual record: the Presi%vztial Transcripts. This comparison will enable us to assss the accuracy of rlis memory rather, precisely. In addition, it may clarify our theoretical conceptions of memory itself. When Dean first testified, his “facility for recalling de&ails” seemed so impressive that some writers called him “‘the human tape recorder”. Ironically, a vey real tape recorder had been tuned in to some of the same “details”. Not long after its interrogation of Dean, the Senate Committee discovered that all conversations in Nixon’s Oval Office were routinely (but secretly) recorded. The result of this discovery was a sharp legal struggle for possession of the tapes. When the President realized that he would not be able to keep the tapes out of the hands of the prosecutors indefinitely, he decided to transcribe some of #em and release the transcripts himself. Although he did this relucrantly, be also thought it possible that they might actually help his cause. The published vtrsion of the f%=esickntialTranscripts (1974 : includes a lengthy foreword reiterating Nixon’s claim that he knew nothing of the cover-up. (It does admit that there are ‘*...possible ambiguities that... someone with amotive to discredit the President could take out of context and distort to suit his own purposes,” - p. 5.) T!!e foreword explicitly insists that the transcripts discredit Dean’s testimony, Dean himself, however, saw them as substantiating his side of the story. In his autobiography (Dean, 1976) he describes himself as “ecstatic” (p. 332) to learn of the tapes’ existence, because they would prove he had told the truth. The testimony and the transcripts am now in the public domain. I propose to treat them as data, as if they had resulted from a deliberately conducted memory experiment. The analysis of these data will be somewhat unorthodox, however, because we know its outcome in advance. If Dean had actually perjured himself - if the transcripts had proved him to be fundamentally mistaken or dishonest - the defense lawyers in the subsequent Watergate trials would surely have seized the opportunity to discredit his testimony. Instead, the outcome of those trials has vindicated him: the highest-placed members of the White House staff all went to prison for doing what John

John Dean’s memory: A case shrdy

3

Dean said they had done. Nixon, of course, was forced to resign. If history has ever proven anything, it surely proves that Dean remembered those conversations and told the truth about them. I will not quarrel with that assessment here, but we shall see that “truth”, “accuracy”, and “memory” are not simple notions. Dean’s testimony was by no means always accurate. Yet even when he was wrong, there was a sense in which he was telling the truth; even when he was right, it was not necessarily because he remembered a iearticular conversation well. These are levels of analysis with which psychology has rarely been concerned. Although there have been many demonstrations of the fallibility of testimony (Stern, 1904; Buckhout, 1974), none hds dealt with a situation as co;nplex as Dean’s: with such significant material, such long spans of time, or such ambiguous motives. We will find it hard to do full justice to John Dean’s memory within the conceptual framework of the psychology of memory. Nevertheless, that framework is not irrelevant. It includes a number of valuable ideas: that memory is influenced by mental “scripts” or “schemata” fol, familiar events (Bartlett, 1932; Bransford & Franks, 1972; Bower, Black & Turner, 1979); that distortions of memory are often motivated by the needs and character of the individual (Freud, 1899); and that a person’s general knowledge (“‘semantic memory”) must be distinguished from his recollection of specific events (“episodic memory”, Tulving, 1972). Most obviously, we will have to make a distinction that has been L;;i;;A at least since Bartlett: to contrast verbatim recall with memo-y for the gist of what was said. Verbatim recall is word-for-word reproduction. It is not something that we expect of ourselves in everyday life. Dean did not claim to be able to recall conversations verbatim, and indeed he could not. ,We shall see that even the few phrases that he seemed to recall exactly may owe their fidelity to frequent repetition.; Memory for gist, on the other hand, occurs when we recall the “sense” of an original text in different words. To remember the gist of a story or a conversation is to be roughly faithful to the argument, the story line, the underlying sequence of ideas. Psychologis’s have developed a number of methods of evaluating memory for gist. One can divide the text and the recall protocol into so-called “idea units”, and count how many of them match. With somewhat more trouble. one can make a structural analysis of the original, perhaps guided by theoretical ideas about “story grammars” and “schemata”; then one can determine how much of the structure reappears in the reproduction (e.g., Mandler & Johnson, 1977). These methods have worked well in the laboratory, where there is nothing to remember except an originally presented text. They are not as easily applied to the recall of actual conversations that take place in a context of real events: The events may be remembered even when the gist of the conversation is not.

4

Utric Neider

Analysis of Dean’s testimony does indeed reveal some instances of memory for the gist of what was said on a particular occasion. Elsewhere in his testimony, however, there is surprisingly little correspondence between the course of a conversation and his account of it. Even in those cases, however, there is usually a deeper level at which he is right. He gave an accurate portrayal of the real situation, of the actual characters and commitments of the people he knew, and of the events that lay behind the conversations he was trying to remember. Psychology is unaccustomed to analyzing the truthfulness of memory at this level, because we us~uallywork with laboratory material that has no reference beyond itself. One of my purposes in analyzing John Dean’s t&imony is to call attention to this level of memory,-and perhaps to devise ways in which it can be studied. Dean%own account of his memory It is impossible to survey all Df Dean’s testhnony here; there is far too much of it. Moreover, most of his converstions were not recorded at all (so far as we know); it was only in the Presideat’s Oval Office that ta.pe recorders ran night and day. Not even ah of the taped materiai :s My reproduced in the available transcripts. We will only be able to analyze the two conversations reported in his testimony for which an apparently unedited transcript has been published. Thereadershould bear in mind that we are dealing with only a small fraction of what Dean said. The present paper is not an effort to assess his overall contribution to the Watergate investigations or to the course of justice; it is a psychological study aimed at clarifying the nature of memory for conversations. The two convetxations we will ex&$ne are those of September 15, 1972 and March 21, 1973. These two meetings with the President were crucial for the Senate Committee, which was trying to determine the extent of Nixon’s involvement in the Watergate cover-up. Accord.tigly, Dean was cross-examined about both of them at length. He had already described each conversation in his long opening statement to the Committee: it was that statement which aroused &ator houye’s incredu&y. The interchange between Deti and Jnouye is interesting irr its own right: it may be the only discussion of mnemonics and metamemory in the Congressional Record. &natur 1.e Your 245page statement is remarkablefcr the detail with which it recruits even*s and conversationsoccurring over a period of many months. It is partictirly remarkabbiu view of the fact that you indicated that it was prepared wihuti benefit of note or daily diary. Would you describewhat documents were availableto you in addition to those which have been identified as exhibits?

John Dean S memory: A case study

5

Mr. Dean What I did in preparing this staten it, 1 had kept a newspaper clipping file from roughly June 17 [June 17, 1972 was the date of the Walergare ureak-in], up until about the time these hearings started when I stopped doing any clipping with any regularity. It was by going through every single newspaper article -utlining what had happened and then placing myself in what I had dtine in a given sequence in time, 2:was aware of all the principal activities I had been involved in, the dealings I had with others in relationship to those activities. Many times things were in response to press activities or press stories that would result in further activities, I had a good mt‘mory of most of the highlights of things that had occurred, and it was through this process, and being extremely careful in my recollection, particularly of the meetings with the President (Hearings,pp. 1432-1433).

Note that Dean has spontaneously invented the temporal equivalent of an ancient mnemonic device: the famotis “method oF loci”. In that method, one mentally moves through a familiar series of places in order to recall images that were previously assigned to them. Dean apparently used newspaper clippings in a similar way, to pinpoint moments in time rather than loci in space; then he tried to recall what he had been doing at those moments. Senator Inouye’s next questions (I am omitting some additional comments by Dean) indicate that he failed to grasp this point; Senator Inouye Are you suggesting that your testimony was primarily based upon

press accounts? !‘rir.Dean No sir, I am saying that I used the press accounts as one of the means to trigger my recollection of wbar bad occurred during given periods of time.

lnouye still does not understand: Senator Inouye Am I to gather from this that you had great fti”a in the reporting in the press? Mr. Dean No, I am saying what was happening is that i.nissequentially - many times White House activities related to a response to a given press activity. I did no: have the benefit - in fact, the statement might be even more detailed, Senator, ii1 had had the benefit of all the Ziegler briefmgs where some of these questions came up very specifically in press briefings as to given events at that time, but I didn’t have the benefit of those (Ibid.). Senator Inoye In addition to the press clippings, the logs, what other sources did you use in the process of reconstruction? Mr. Dean Well Senator, I think I have a good memory. I think that anyone who recalls my student years knew that I was very fast at recalling information, retaining information. I was th’etype of student who didn’t have’to work very hard in school because I do havea memory that I think is good (Ibid.). A moment later Inouye asks the question I have already quoted, encouraging Dean to say more about his memory:

Senotar Inouye Have you always had a facility for recallingthe detai%of conversations which took place many months ago? (Bid.). Dean responds with examples of thmgs he would certainly never forget,

beginning with conversations in the Oval Office: MCDemrWell,I wolnldlike to start with the Presidentof the United States. It was not a regularactivity for me to go in and visit with the President.FCM most of the membersof the White House staff it was not a daily activity.Whenyou mee; with the Presidentof the United States it is a very momentousoccasion,and you tend to

rememberwhat the l%esidentof the United States says when you have a conversation with him. (Deangoeson to mentiv>nseveralother salientevents that he remembers ~11, and conckdes...) 230 I would say that I have an ability to recall not specificwordsmzcessarilybut certainlythe tenor of a conversationand the gist of a conversation(Ibid, pp. 1433-1434). We shall see later that Dean recalls the “g&t” of some conversations and not of others; the determinants of memory are more complicated than he believes them to be. In part&&r, he did not remember what the President said in their fmt prolonged and “momentous” meeting. But there is no doubt about his contidenceinhir, own testimony: at the end of the exchange with Inouye, he expfesses it again: Mr. L&n 1 cannot repeal the very words he (the President)used, no, sir. As I explained to Senatol Gurney, my mind is not a tape recorder,but it certainlyreceives the messagethat is-beinggiven (fbti ).

The Meeting

of September 15

On June 17, 1972, dive men were arrested in the offices of the Democratic NationaI Committee in the Watergate Office Building. They had planned to tap the Committee’s telephones as part of an illegal *‘politicaI intelhgence” operation, mounted on Resident Nixon’s behalf in the I972 presidential elections. High White House officials then began a major effort to conceal their lnvoIvement in the affair, even to the point of paying “hush money” to some of those who had been arrested. John Dean was centrally involved in the cover-up. His chief task was to “contair? thc9legal investigation of the Watergate break-in, concealing every link between the underlings already caught and the White House. Qn September 15 this aim seemed achieved, because on that day the Grand Jury handed down indictments against only seven men: the five burglars plus Howard Hunt and Gordon Liddy. Since

John Dean’s memory: A cllse study

7

Hunt and Liddy were “small fish”, and the Justice Department said it had no evidence to indict anyone else, Dean felt victorious. When the President summoned him to the Oval Office that afternoon, he expected to be praised. The transcript indicates that the meeting lasted 50 minutes. It begins with the following interchange among the President (P), Dean (D), and Robert Haldeman (H), Nixon’s “Chief of Staff”. Note that Dean and Haldeman are both obviously pleased by the events of the day, while the President has little to say about them. P Hi, how are you? You had quite a day today, didn’t you? You got Watergate on the way, didn’t you? D We tried. I” How did it all end up? D Ah, I think we can say well, at this point. The press is playing it just as we expected. H Whitewash? D No, not yet - the story right now P It is a big story. H Five indicted plus the WH former guy and all that. D Plus two White House fellows. H That is good; that takes the edge off the whitewash, really. That was the thing Mitchell kept saying, that to people in the country Liddy and Hunt were big men. Maybe that is good. P How didMacGregor handle himself? D I think very well. He had a good statement, which said that the Grand Jury had met and that it was now time to realize that some apologies may be due. H Fat chance. D Get the damn (inaudible) H We can’t do that. P Just remember, all the trouble we’re taking, we’ll have a chance to get back one day. How are you doing on your other investigation? (Presidentialllanscripts, p. 32)

The next few exchanges are about other details of the Watergate “bugs” (telephone taps), and then about the scope of the investigations being conducted. It all seemed %lly” to them, especially since they believed that “bugging” was common in politics: P Yes (expletive deleted). Goldwater put it in context when he said “(expletive

deleted) everybody bugs everybody else. You know that”. D That was priceless. P It happens to be totally true. We were bugged in ‘68 on the plane and even in ‘62 running for Governor -- (expletive deleted) thing you ever saw. D It is a shame that evidence to the fact that that happened in ‘68 was never around. I understand that only *the former director (J. Edgar Hoover, former head of the FBI) had that inform&on.

8

IJZricNeisser

H No, that is not true. D There wasevidenceof it? H There are others who have information(Ibid.

p. 34).

This interchange about “bugging”is noteworthy :gQtonly because of the light it sheds on the attitudes of the participants, ;r:ut also because it stuck in Dean’s mind, It is one of the few parts of the conversation wrlich will be recognizable in his testimony nine months later. The conversation cont!nues from this point with more talk about “‘bugging”, plans for action against White House enemies, questions about another pending legal action. It is interrupted briefly when Nixon takes a phone call. As soon as he hangs up, Dean spe:rks. He wants to point out how well things are going: .D Three months ago I would have had trouble predicting there would be a day when this would be forgotten, but I think I can say that 54 days from cow [i.e., on election day in November! nothing is going to come crashing down to our surprise. Y That what? D Nothing is going to come crashing down to our surprise (Wd, p. 36).

He finally gets a bit of Presidential praise in return: P Oh well, this is a can of worms as you know 3 lot of this stuff that went on. And

the people who worked this way are awfully embarrassed. But the way you have handled all this seems to me has been very skJlfu1, putting your fingers in the leaks that have sprung here and sprung there. The Grand Jury is dismissed now? D That is correct... (Ibid.).

The conversation goes on to cover many other areas - McGovern’s campaign finances, a list of “enemies” that Dean offers to keep, more political strategy. Later on Dean and Haldeman (but not Nixon) seize another opportunity to congratulate each other on the success of the cover-up. P You really can’t sit and worry about it all the time. The worst may happen but m;v not. So you just try to butto:s it up as well as you can and hope for the best, and remember basically the damu business is unfortunately trying to cut our losses. D Certainly that is right and certainly it has had no effect on you. That’s the good t1hing. H BJo, it has been kept away from the White House and of course completely from the President. The only tie to the White House is the Colson effort thei keep trying to pull .in. D And of course the two White House people of lower level - indicated .- one consultant and one member of the do.nestic staff. That is not very much of a tie. H That’s right (Ibid, p_ 40).

John Dean b memory: A case study

9

Dean’s testimony about September 15 Nine months later, Dean devoted about two pages of his prepared statement to the September 15 meeting. The first paragraph purports to describe the way the meeting began. It is an important bit of testimony because the remarks Dean ascribes to Nixon would indicate full knowledge (and approval) of the cover-up. This is his account: On September15 the Justice Department announced the handing down of the seven indictments by the Federal Grand Jury investigating the Watergate. Late that afternoon I received a call requesting me to come to the President’s Oval Office. When I arrived at the Oval Office I found Haldeman and the President. The President asked me to sit down. Both men appeared to be ii1very good spirits and my reception was very warm and cordial. The President then told me that Bob - referring to Haldeman - had kept him posted on my handling of the Watergate case. The President told me I had done a good job and he appreciated how difficult a task it had been and the President was pleased that the case had stopped with Liddy. I responded thea I could not take credit because others had done much more”difficult things than I had done. As the President discussed the present status of the situation I told him that all I had been able to do was to contain thp case and assist in keeping it out of the White House. I also told him there was a long way to go before this matter would end and that I certainly could make no assurances that the day would not come when this matter would start to unravel (Hearings,p. 957). Comparison

with the transcript.

shows that hardly

a word of Dean’s ac-

count is true. Nixon did not say asiy of the things attributed to him here: he didn’t ask Dean to sit down, he didn’t say Haldeman had kept him posted, he didn’t say Dean had done a good job (at least not in that part of the conversation), he didn’t say anything about Liddy or the indictments. Nor had Dean himself said ths things he later describes himself as saying: that he couldn’t take credit, that the matter might unravel some day, etc. (Indeed, he said just the opposite later on: “nothing is going to come crashing down”.) His account is plausible, but entirely incorrect. Jn this early part of the conversation Nixon did not offer him any praise at all, unless “You had q&e a day, didn’t you”, was intended as a compliment. (It is hard to tell from a written transcript.) Dean cannot be said to have reported the “gist” of the opening remarks; no count of idea units or comparison of structure would produce a score mu& above zero. Was he simply lyi.ng to the Senators? I do not think so. The transcript makes it quite clear that Nixon is fully aware of the cover-up: Haldeman and Dean discuss it freely in front of him, and while he occasionally asks questions he never seems surprised. Later on he even praises Dean for “putting his fingers in the leaks”, Because the real conversation is just as incriminating

10

Ulric Neisser

as the one Dean described, it seems unlikely that he was remembering one thing and saying another. His responses to Senator Baker during crosse unination (see below) also indicate that he was doing his best to be honest. Mary McCarthy’s assessment of Dean has stood the test of time: she wrote in 1973 of her overpowering impression “ . ..not so much of a truthful person as of someone resolved to tell the truth about this particular set of events because his intelligence has warned him to do so” (McCarthy, 1975, op. 40-4 1). If Dean was trying to tell the truth, where did his erroneous account of the September 15 meeting come from? Some of it might be explained by the currently popular notion that everyone knows certain “scripts” for common events, and that these scripts are used in the course of recall (Bower, Black, and Turner, 1979). Dean’s recollection of the very beginning of the meeting may have been constructed on the basis of an “entering-the-room script”. People do often ask their guests to sit down, though Nixon apparently did not ask Dean. It is also possible, however, that Dean’s recollection of such a request is a case of non-verbal gist recaIl rather than a script-based construction. Perbaps Nixon did ask Dean to sit down, but with a gesture rather than a word - a brief wave of a commanding presidential han.d. To recall such a gesture as if it had been a verbal request wou-Id not be much of an error. Current theoretical interest in the recall of written texts should not blind us to the non-verbal components of real conversation. Although familiar scripts and non-verbal cues explain a few of Dean’s errors, most of them seem to have deeper roots. They follow, I believe, from Dean’s own character and especially from his self-centered assessment of events at the White House. What hrs testimony really describes is not the September 15 meeting itself but his fantasy of it: the meeting as it should have been, so to speak. In his mind Nixon should have been glad that the indictments stopped with Liddy, Haldeman should have been telling Nixon what a great job Dean was doing; most of all, praising him should have ileen the fast order of butsiness. In addition, Dean should have told Nixon that the cover-up might unravel, as it eventually did, instead of telling him it was a great success. By June, this fantasy had become the way Dean remembered the meeting. Almost. But Dean was not really as confident of his recollection as the tone of his statement suryested; not as sure of himself as he claimed in the exch;urge with Senator Inouye. This becomes clear in a very sharp interrogation by Senator Baker: Senator Baker I am goingto try now to focus entirely on the meetingof September 15. Mr. Dean Right.

John Dean’s memory: A case study

11

Senator Baker And I have an ambition to focus sharply on it in order to disclose as

much information as possible about the September 15 meeting. What I want to do is to test, once again, not the credibility of your testimony but the quality of the evidence, that is, is it direct evidence. Mr. Dean I understand (Hearings, p. 1474).

Dean does understand: Baker wants vivid details and exact wording. The next few exchanges show how he struggles to reconcile the vagueness of his actual recollection with Baker’s demands for specificity, dodging some questions and eventually committing himself on others. After an uncontroversial account of how he learned that Nixcu wanted to see him that evening, Dean begins with his physical entrance into the office: Mr. Dean

WhenI entered the office I can recall that - you have been in the office. you know the way there are two chairs at the side of the Presideat’s desk. Senator Baker You are speaking of the Oval Office? Mr. Dean Of the Oval Office. As you face the President, on the left-hand chair Mr. Haldeman was sitting and they had obviously been immersed in a conversation and the President asked me to come in and I stood there for a moment. He said “Sit down”, and I sat in the chair on the other side. Senator Baker You sat in the right-hand chair? Mr. Dean I sat on t:,e right-hand chair. Senator Baker That is ‘the one he usually says no to, but go ahead. Mr. Dean I was unaware of that. (Laughter). Senator Baker Go ahead, Mr. Dean (Ibid, p. 1475). Now Dean plunges into the conversation, giving almost exactly the same account of it that he had presented in his prepared statement a few days before. Indeed, his opening phrase suggests that he is remembering that statement rather than the meeting itself: Mr. Dean As I tried to describe in my statement, the reception was very warm and

cordial. There was some preliminary pleasantries, and then the next thing that I recall the President very clearly sayiclgto me is that he had been told by Mr. Haldeman that he had been kept posted or made aware of my handling of the various aspects of the Watergate case and the fact that the case, you know, the indictments had now been handed down, no one in the White House had. been indicted, they had stopped at Liddy (Ibid.).

Senator Baker is not satisfied with this response; he wants to know how accurate Dean is really claiming to be: Senator Baker Stop, stop, stop just for one second. “That no one in the White House

had been indicted”: is that as near to the exact language - I don’t know so I am not laying a trap for you, I just want to know (Ibid.).

12

LCc Neisser

It is now clear that the right answer to Baker’squestion would have been “no”. Nixon did not use anything remotely like the “exact language” in question; the conversation did not go that way at all. Dean’s answer is cautious: Mr. Dean Yes, there was a reference to the fact that the indictments had been

handed down and it was quite obvious that no one in the White House had been indicted on the indictments that had been handed down (Zbid.). Notice that although Dean’s answer begins with “Yes”, he no?4 avoids attriI-sting the critical words to Nixon. He hides behind ambiguol;s phrases like “There was a reference to the fact that...” and “It was quite obvious...” Baker is unsatisfied with these evasions and continues to press for a straight answer: Senator Baker Did he say that, though? (Ibid.).

Dean decides to be honest about it: Mr. Dean Did he say chat no one in the white House had been ha&d can’t recall it. (Ibid.).

down? I

This is the answer which suggests to me that Dean was being as truthful as he could. After all, he might easily have answered “yes” instead of “I can’t recall it”. But he doesn’t want to give up the points he has already scored, so he repeats them: Mr. Dean (continuing) J can recall a reference to the fact that the indictments were

now handed doiNn ar;d he was aware of that and the status of the indictments and expressed what t? me was a pleasure to the fact that it had stopped with Mr. L+iddy (Zbid.). This paragraph versation, and it was reference to of that (though

is a nice sumunary of what Dean remembers from the conis phrased so carefully that everything in it is true. There the indictments (by Haldeman and Dean); Nixon wus aware he didn’t say so); and somehow he did express what Dean

interpreted as pleasure in the outcome. It is fair to say that Dean here captures fthe “tenor”, though not the gist, of what went on in the Oval Office that afternoon. But Baker notices that he still hasn’t committed himself to any exact statements by Nixon, and tries again: Senator Buker Tell mc what he said. Mr Dean WeJl,as I say, he told me I had done a good job Senator Baker No, let’s talk about the pleasure. He expressed pleasure the indict. ments had stopped at Mr. Liddy. Can you just for the purposes of our information tell me the languatiethat he used? Obid.).

Dean ducks once more:

John Dean’s memory: A caxe study

13

Mr. Dean Senator, let me make it very clear: the pleasure that it had stopped there is an inference of mine based on, as I told Senator Gurney yesterday, the impression I had as a result of the, of his, complimenting me (J&-L).

Baker hangs tough: Senator Baker Can you give us any information, can you give us any further insight

into what the President said? Mr. Dean Yes, I can recall he told me that he appreciated how difficult a job it had been for me. Senator Baker Is that close to the exact language? Mr. Dean Yes, that is close to the exact language (Ibid., p. 1476).

Finally Dean gives in, and puts words into Nixon’s mouth. He may just have felt he had no choice: if he didn’t claim to remember any of Nixon’s remarks his whole testimony might be discredited. But also he may have believed it. Nixon’s compliment was what he had most yearned for, and his invented version of it may have been the most compelling thing in his memory. Either way, the exchange seems to have hardened his willingness to testify to exact language. He and Baker went at it again a few minutes later when Dean said he had told Nixon “that the matter had been contained”. Baker repeatedly asked whether he had used that very word, and Dean repeatedly asserted that he had done so. When Baker questioned him closely about how the President had reacted to “contained”, however, Dean said he did not recall. He certainly didn’t: the word “contained” appears nowhere in the transcript. In summary, it is clear that Dean’s account of the opening of the September 15 conversation is wrong both as to the words used and their gist. Moreover, crossexamination did not reveal his errors as clearly as one might have hoped. The effect of Baker’s hard questioning was mixed. Although it did show up the weakness of Dean’s verbatim recall, ii:,. o--srall rcsialt may have been to increase his credibility. Dean came across as a man who has a g~cl~:: memory for gist with an occasional literal word stuck in, hke a raisin in a pudding, He was not such a man. He remembered how he had felt himself and what he had wanted, together with the general state of affairs; he didn’t remember what nnyone had actually said. His testimony had much truth in it, but not at the level of “gist”. It was true at a deeper level. Nixon was the kind of man Dean described, he had the knowledge Dean attributed to him, there was a cover-up. Dean remembered all of that; he just didn’t recall the actual conversation he was testifying about. So far I have concentrated on the first few minutes of the meeting, covered in a single paragraph of Dean’s prepared statement. The next paragraph is

interesting because (unlike the i’bst) it refers to a bit of conversation that

actually occurred. Easfyin ourconversationthe Presidentsaid to me that formerFBI DirectorHoover had told him shortly after he assumed office in 1969 that his campaignhad been bugged in 196%The Presidentsaid that at some point we should get the facts out on this and use this tcr counter tl =problemsthatwe were encountering(Ibid. p. 958). AS we have already seen, an exchange about Hoover and bugging in previous campaigns did take place, a little after the beginning of the conversa-

tion: But although it was indeed Nixon who raised the subject, it was Dean, not Nixon, who brought Hoover’~ name into it: “I understand that only the former director had that information”. Dean may have forgotten this becattse Haldeman had put him Gown so &arply (“No, that is not true”), or he may have preferred to put the words into Nixon’s mouth for other reasons. In any case, he isn’t quite right. The remainder of Dean’s tes-timory about the meeting is no better than the parts we have examired. He mentions topics that were indeed discussed, but never reproduce, the real gist of anything that was said. Surprisingly, he not remember the President’s actual compliment to him (“putting your fingers in the leaks”) although it is a fairly striking phrase. At the end of his statement he presents the followklg summary: MYD~XBI “I left the meeting with the impresxon that the Presidentwas well aware of what had been going on regardingthe succe~*of keeping the White House out of the Watergatese&al, and I also had expressedto him my concern that I was not confident that the cover-upcould be maintainedind&litely (Ibid., p. 959). The first part of this summary is fair enough: Nixon was surely “...weJl aware of what had been going on”. The conclusion is less fair; Dean seriously - perhaps deliberately - misrepresents the optimistic predictions he ha:1

made. In fact he was nut wise enough or brave enough to warn Nixon in September, though by June hs was smart enough to wish he had done so.

The cover-up was only temwrarily successful. Although N&on was reelected ovenvhehningly in November of 1972, Dean’s problems increased steadily. There were more blackma demands by the indicted Watergate defendants, and more investigations moving closer to the White House. Dean met frequently with Nixon, Haldeman, and the others, but their strategems were unsuccessful. Dean began to realize that he and the others were engaging in a

John Dean’s memory: A case study

15

crime (“obstruction of justice”), and might eventually go to prison for it. He was not sure whether Nixon understood the gravity of the situation. Finally he resolved to ask the President for a private meeting at which he could lay out all the facts. This meeting took place on March 2 1, 1973. Dean’s autobiography (1976) relates an incident that occurred on the day before the critical meeting. When he was trying to describe the relentlessly ilrcreasing complexity of the Watergate affair to Richard Moore, another White House aide, Moore compared it, to the growth of a tumor. The metaphor attracted Dean, and he resolved to use it in his report the next day: to tell Nixon that there was a “cancer” growing on the presidency. The transcript of the meeting shows that he did so. After a few minutes of conversation about the day’s events, Dean and the President continue as follows: D The reason I thought we ought to talk this morning is because in our conversations I have the impression that you don’t know everything I know, and it makes it very difficult for you to make judgments that only you can make on some of these targets, and I thought that P In other words, I have to know why you feel that we shouldn’t unravel something? D Let me give you my overall first. P In other words,your judgment as to where it stands, and where we will go. D I think there is no doubt about the seriousness of the problem we’ve got. We have a cancer within, close to the presidency, that is growing. It is growing daily. It’s compounded, growing geometrically now because it compounds itself. That *willbe clear if I, you know, explain some of the details of why it is. Basically it is because (1) we are being blackmaited; (2) people are going to start perjuring themselves very quickly that have not had to perjure themselves to protect other people in the line. And there is no assurance P That that won’t bust? D That that won’t bust (PresidentialTranscripts,pp. 98-99). In this fmt part of the March 21 meeting, Dean was alone with the President. They remained alone for about an hour, then Haldeman came in *o join the discussion for another 45 minutes or so. Haldeman’s entrance proved to be a critical turning point in Dean’s later memory of that morning: he forgot the rest of the conversation almost completely. What he said abo’ut the fmt hour, in contrast, was quite accurate. Comparison of tile transcript with Dean’s subsequent testimony shows clear recall of the gist of what was said. One’s admiration for his memory is somewhat diminished, however, by the realization that the March 21 meeting was less a conversation than the delivery of a well-prepared report. Dean did most of the talking, taking 20 minutes to describe the events before the break-in and 40 more for the coverup. Although Nixon inteejected occasional remarks, questions, or expletives, the hour stayed quite do& to the script Dean had prepared for it in advance.

The difference between this meeting and that. of September 15 is instructive. This one fulfiied Dean’s hopes as the earlier one had not: he really did give a personal lecture to the President of the United States, talking while Nixon listened. His testimony, too long to reproduce here, bighlights the meeting’s didactic quality. Almost every statement begins with “I told him...“, “I proceeded to tell him...“, “I informed the President...” or some similar phrase. He was remembering a report that he had rehearsed ahead of time, presented as planned, and probably continued to rehearse aifterwards. It bepae John Dean’s own story; March 21 had merely been his fiit opportunity to tell it. Dean’s testimony includes a fragment of nearQ verbatim recall that later achieved some notoriety: he quoted his own remark about the “cancer on the presidency” to the Senate Committee. This, too, was a well-rehearsed passage. We know that he prepared it in advance, and the transcript shows that he used it repeatedly. (He probably used it on other occasions as well; why let such a good phrase go to waste?) His frost presentation of the simile, early in the meeting, has been quoted above. Twenty minutes later he refers back to it: D ...When1saythis is a growingcancer,I say it for reasonslike this..,,(Ibid., p. 111). and still later he brings it in obli+ely: D ...we should begin to think... how to minimize the furthergrowth of this thing... (Ibid., p. 119).

Interestingly, Dean’s selfquotation of these occasions:

to tk Senators was not faithful to any

I began by telling the President that there was a cancer growingon the presidency and Ehatif the cancerwas not removedthe Presidenthimself wrouldbe killed by it. I ak~ told him that it was imprtant that this cancer be removed immediately becauseit was growingmote deadly everyday (Hemings, p. 998). A glance back at the excerpt from the transcript shows that Dean is once again giving himself the benefit of hindsight. He did not say that the President wouid be killed by the cancer, for example. By June he probably wished he had done so; I don’t know whether he altered the wording in his testimony deliberately or whether his memory had already accommodated itself slightly to his self-image. in Dean’s mind, the signi&ance of the March 21 meeting must have lain in the degree to which he -dominated it. That may explain why he barely mentioned the second half of the meeting in his Senate testimony; Haldeman’s entrance spoiled his private command performance. The rest of the session

John Dean ‘smemory: A case study

17

was by no means uninteresting, however. What actually happened was that Nixon, Haldeman, and Dean considered various options, trying to find the best way to deal with their Watergate dilemma. One of those options was to raise money to meet the blackmail demands of ths men who had already been convicted. This possibility seemed to attract Nixon; he returned to it again and again. He had aheady discussed it in the first hour, when only Dean was with him : D I would say these people are going to cost a million dollars over the next two years. P We could get that. On the money, if you need the money you could get that. You could get a million dollars. You could get it in cash. I know where it could be gotten. It is not easy but it could be done... (I+esidential Transcripts,p. 110).

He seemed more enthusiastic about it than Dean himself: P Just looking at

the immediateproblem,don’t you think you have to handle Hunt’s financial situation damn soon? D I think that is - I talked with Mitchell about that last night and P It seems to me we have to keep the cap on the bottle that much or we don’t have any options (Ibid., p. 112). Later he makes it as explicit as he possibly can: D The blackmailers. Right. P Well I wonder if that part of it can’t be - I wonder if that doesn’t - let me put it frankly: I wonder if that doesn’t have to be continued? Let me put it this way: let us suppose you get the million bucks, and you get the proper way to handle it. You could hold that side? D Uh-huh. Pit would seem to me that would be,wo’rthwhile(Ibid., p. 117).

Remarks like them:

this

continue to sprinkle the conversation after Haldeman joins

P ...First. it is going to require approximately a million dollars to take care of the jackasses who are in jail. That can be arranged... (Ibid., p. 127). ... P Now let me tell you. We could pet the money. There is no problem in that... (Ibid., p. 129).

... P I just

have a feelingon it. Well, it soundslike a lot of money a million dollars. Let me say that I think we could get that... (Ibid., p. 130).’

‘Nixon never expressed qny hesitation about mtig these payments, or any reluctance to meet the burglars’ demands for money. He did, however, agree with Dean that their demands for executive clemency should not be met. At one po,int he said “NO - it is wrong, that’s for sure” about the possibility of clemency. The transcript shows no analogous statement about the blackmail payments.

18

Ulric Neisser

‘IJ’heseare quite remarkable things for a President to say. They would certainly seem to be memorable, and indeed Dean did not forget them. He just assiged them to a different day! Although he makes no reference to them in Thistestimony about March 2 1, his statement includes the following description of a meeting with Nixon on March 13, eight days before: .A was during this conversation *hat Haldeman came into the office. After this brief interruptionby Haldeman’scoming in, but while he was still there, I told the Presidentabout the frsctthat there was no money to pay these individualsto meet their demands.He asked me how much it would cost. I told him that I could only make an estimatethat it might be as Qgh as $1 million or more. He told me that that was no problem, and he also looked over at Haldemanand made the same statement...(Hearings,p. 995). Dean amplifies this ticcount latfp -

%tg crossexamination:

...Wehad also had a discussion on March13 about the money demands that were b&g made, At the time he discussedthe fact that a million dollarsis no problem. He repeatedit severaltimes. I can very vividly recall that the way he sort of rolled his chair back from his desk and leaned over to Mr.Haldemanand said “A million dollarsis no problem”(Ibid., p. 1423).

It is hardly surprising that Dean remembered these milliondollar statements, especially since Nixon repeated them so often. It is a little surprising that he put them into the wrong conversation. (There is a transcript of the Mach 13 meeting, and it shows no such remarks by the President.) Evidently ajean’s improvised method of temporal loci, based on newspaper clippings, cud not work as well as his exchange with Senator Inouye had suggested. His ego got in the way again. The March 21 meeting had been the occasion for his CWIIpersonal report to the President; he could not suppose that anything else worth mentioning had happened. Other memories were shifted to another day if they survived at all. Nixon’s eagerness to pay the blackmail money was not the only part of the conversation to suffer this fate. Dean even displaced one of his own jokes; a joke that had drawn a response from Haldeman if not from Nixon. They were discussing various illegal ways of “laundering” the blackmail money so it could not be traced: D And that means you have to go to Vegas with it or a bookmakerin New York City. I have learnedall these &ings after the fact. I will be in great shape for the next time around! H (Expletivedeleted)(&Wder&l lbnmipl~, p. 134).

John Dean’s memory: A case st’udy

19

That may not have bee% the only time Dean used this wisecrack; he probably enjoyed describing himself as increasingly skilled in underworld techniques. Certainly he’didn’t mind repeating it to the Senators, though his statement assigns it, too, to March 13 rather than March 21: . ..I told him I was learning about things I had never sad before, but the next time I would certainly be more knowledgeable. This comme;lt got a laugh out of Haldeman (Hearings, p. 996).

It isn’t very funny.

Implications for the Psychology of Memory Are we all like this? Is everyone’s memory constructed, staged, self-centered? And do we all have access to certain invariant facts nevertheless? Such qdestions cannot be answered by single case h&tories. My own guess - and it is only a guess - is that reconstruction played an exaggerated part in Dean’s testimony. The circumstances and the man conspired to favor exaggeration. The events were important; his testimony wetscritical; its effect was historic. Dean was too intelligent not to know what he was doing, and too ambitious and egocentric to remain unaffected by it. His ambition reorganized his recollections: even when he tries to tell the truth, he can’t help emphasizing his own role in every event. A different man in the same position might have observed more dispasionately, reflected on his experiences more thoughtfu9y, and reported them more accurately. UnfiJrtunately, such traits of character are rare. What have we learned about ‘testimony by comparing “the human tape recorder” with a real one? We are hardly surprised to find that memory is constructive, or that confident witnesses may be wrong. William Stem studied the psychology of testimony at the turn of the century and warned us not to trust memory even under oath; Bartlett was doing experiments on “constructive” memory fifty years ago. I believe, however, that John Dean’s testimony can do more than remind us of their work. For one thing, his constructed memories were not altogether wrong. On the contrary, there is a sense in which he was altogether right; a level at which he was telling the truth about the Nixon White House. And sometimes - as in his testimony about March 2 1 - he was more specifically right as well, These islands of accuracy deserve special consideration. What kinds of things did he remember? Dean’s task as he testified before the Senate Committee was to recall specific well-defined conversations, “...conversations which took place months ago”. This is what wit,zesses are always instructed to do: stick to the

20

UhicNeisser

facts, avoid inferences and generalizations. Such recall is what Tulving (1972) called episodic; it involves the retrieval of particular autobiographical moments, individual episodes of one’s life. Tulving contrasted episodic memory only with what he called semantic memory, the individual% a,gcumulated store of facts and word meanings and general knowledge. That concept seems inadequate as ,:1description of data such as these. Dean’s recollection of Nixon’s remark> about the million dollars was not merely semantic: he talked as if he were rcl:alling one or more specific events. I doubt, however, that any of thos: events was being recalled uniquely in its own right. A single such episode might not have found its way into Dean’s testimony at all. What seems to be specific in his memory actually depends on repeated episodes, rehearsed presentations, or overall impressions. He believes that he is recalling one conversation at a time, that his memory is “episodic” in Tulving’s sense, but he is mistaken. He is not alone in making this mistake. I believe that this aspect of Dean’s testimony illustrates a very common process. The single clear memories that we recollect so vividly actually stand for something else; they are “screen memories” a little like those Freud discussed long ago. Often their real basis is a set of repeated experiences, a sequence of related events that the single recollection merely typifies or represents. We are like the subjects of Posner and Keele (1970) who forgot the individual dot patterns of a series but “remembered” the prototypical pattern they had never seen. Such memories might be called repisodic rather than episodic: what seems to be an episode actually represents a repetition. Dean remembers the million-dollar remark because Nixon made it so often; he recalls the “cancer” metaphor because he first planned it and then repeateti it; he remembers his March 21 lecture to the President because he planned it, then presented it, and then no doubt went over it again and again in his own mind. What he says about these “‘repisodes” is essentially correct, even though it is not literally faithful to any one occasion. He is not remembering the “gist” of a single episode by itself, but the common characteristics of a whole series of events. This notion may help us to interpret the paradoxical sense in which Dean was accurate throughout his testimony. Given the numerous errors in his reports of conversations, what did he tell the truth about? 1 think that he extracted the common themes that remaim 1 invariant across many conversations.and many experiences, and then iccokporated those themes in his testimony. His many encounters with Nixon were themselves a kind of “repisode”. There were certain consistent and repeated elements in all those meetings; they had a theme that expressed itself in different ways on different occasions. Nixon wanted the cover-up to succeed; he was pleased when it went well; he was troubled when it began to unravel; he was perfectly willing to

John Dean ‘s memory: A case sudy

21

consider illegal activities if they would extend his power or confound his enemies. John Dean did not misrepresent this theme in his testimony; he just dramatized it. In memory experiments, subjects often recall the gist of a sentence but express it in different words. Dean’s consistency was deeper; he recalled the theme of a whole series of conversations, and expressed it in different events. Nixon hoped that the transcripts would undermine Dean’s testimony by showing that he had been wrong. They did not have this effect because he was wrong only in terms of isola’sed episodes. Episodes are not the only kinds of facts. Except where the significance of his o-,vn role was at stake, Dean was right about what had really been going on in the White House. What he later told the Senators was fairly clase to the mark: his mind was not a tape recorder, but it certainly received the message that was being given.

References Bartlett, F. C., (1932) Remembering. Cambridge University Press. Rower, G. H., Black, J. B., and Turner, T. J. (1979) Scripts in memory for text. Cog. Psychoi., II, 177-220. Bransford, J. D. and Franks, J. J. (4972) The abstraction of linguistic ideas: A review, Cog., I, 211249. Buckhout, R. (1974) Eyewitness testimony. Sci. Amer., 2.31 (6), 23-31. Dean, J. W. (1976) BlindAmbirion. New York, Simon and Schuster. Freud, S. (1899) Screen memories. Reprinted in CollectedPapersof Sigmund Freud, VoI. V. LondoR: Hogarth Press, 1956. Hearingsbefore the Select Committeeon PresidentialCompaignActivities of the United States Senate, Ninety-third Congress, First Session, 1973. Mandler. J. M.and Johnson, N. (1977) Remembrance of things parsed: story structure and recall. Cog. Psychol.. 9, 111-151. McCarthy, M. (1975) 7%e Mask of State: WatergatePortraits.New York, Harcourt Brace Jovanovich (Harvest). Posn~, M.1. and Keele, S. (1970) Retention of abstract ideas: J. &per. Psychal., 83, 304-308. ThePresidential l?anscrlpts (1974). New York, Dell. Stern, W. (1904) Wirklichkeitsversuche (Reality Experiments). Beitrage zur Psychologikder Aussage, 2, No. 1, 1-31. Tub&g, E. (1972) Episodic and semantic memory. In E. Tulving and W. Donaldson (Eds.), Organization and Memory. New York, Academic Press.

John Dean, I’ancien ionseiller du Prisident Nixon, a t& loigne devant le ComitC Sdnatorial sur I’affaire Watergate au sujet dc conversations dont on a constat 21~s tard qu’eUes avaient bt& enregistrkees. La comparalson du timoignage de J. Dean et des transcriptions indique des distortions systgmatiques a UR

22

ufrc Neher

certain niveau d’analyse et tne grande prkision i un autre n&au. De nombreuas distortions sont ii&s i l’ihnage que Dean a dc lui-m&ne, il tend A se rappeler son r6le comme plus central qu’il ne Wait en fait. Par co&e, sa mdmoire pour l’esrentiel des conversations est pauvre sauf dans les cas ou celle-ci a it6 frkquemment rep&e ou s&r&e prklabiement. Alors que son tkmoignage est entacht d’erreun quand il s’agit de conversations partwli6res qu’il es&e de d&ire, Dean est fondamentalement correct 3 propos de ce qui s’est pa&: I’existence d’une “couverture” et la partrcipation de p!usieun personnes 0 celle-ci. Ce t6moignage est dorm p&is 5 un niveau qui n’est ni %mantique” (g&que Dean d&it manifestement des +sodes particuli&res)ini “kpisodique” (puisque son compterendu des 6piies est souvent faux). Le tern?f?‘Iepisodic” a it& c&e pour dicrire de tels souvenirs, il semble qu’un ipisode dont on se souvient consiste en une s&e d’bvenements reflktant un itat de fait awthentique.

Cognition. 9 (1981) 23-33 @ F&evierSequoia S.A.. u8wanne - Printed in the Netherlands

2

RecognitionHypermnesis:The Growth of Recognition Memory (d’) Over Time with RepeatedTesting* MATTHEW HUGH E.qDELYI JUDY B. STEIN

Brook/yn College, C. UN. Y

Abstract This study sought to determine whether recognition hypermnesia (unforgetting) might be obtained over time with repeated testing of recognition memory. Four types of stimuli were investigated: pictures or words (captions) from configured (funny) or non-configured (non-funny) cartoons. Three recognition tests, each comprising the same stimulus and distractor items, were successively administered after the presentation of a large set of captioned cartoons. Silent think intervals of 5 min were interpolated between recognition tests. Recognition hypermnesia, measured by increasing d’ estimates of accessible recognition memory, was obtained with the picfures from con figured (funny) cartoons but not with any of the other groups.

In this article We report the first successful experimental induction Gf recognition hypermnesia. Subjects repeatedly tested for recognition of pictorial components of a set of configured stimuli (funny cartoons) improved their recognition performance (as measured by d’) in successive recognition tests. Such recognition hypermnesia was not obtained with verbal components of configured stimuli (captions of funny cartoons) nor with pictorial or verbal components of noncontigured stimuli (non-funny cartoons). The present work constitutes an extension of recent research on recall hypermnesia by Erdelyi and his associates (Erdelyi and Becker, 1974; Erdelyi and Kleinbard, 1978; Erdelyi, Finlcelstein, Herrell, Miller and Thomas, 1976; Shapiro and Erdelyi, 1974). In the fast of this series of studies, Erdelyi and Becker (1974) demonstrated that recall of pictures (but not words) increased *This mwerch WBSmpg&ted by Public Health Service/National Imtitute of Mental Health Grant 1 RP3 MlH2587641 and by City University of New York Faculty Research Award Fmgram Grants 10704. 11189, 11800, and by Professioaal Staff Congress-Board of Higher Education Grant No. 12216 to the senjor author, Requmta for repzints should be addressed to Matt Erdelyi, Department of Psychology, Brooklyn CoBege,C.U.N.Y., Brooklyn, New York 11210, U.S.A.

24

Matthew ii. Erdelyi and Judy B. Stein

over tjme with successive tests of recall. This effect is both reliable and powerful. Testing subjects for a full week’s duration, Erdelyi and Kleinbard (1978) obtained increments in picture recall averaging in excess of 50% of initial recall levels. The &sic finding has been independently substantiated (e.g., Mad&an, 1976; Yarmey, 1976), though there is some question whether hyrermnesia is necessarily confined to imagistic materials such as pictures or imaged words (Erdelyi et& 1976; Erdelyi, Buschke, and Finkelstein. 1977; Roediger and Thorpe, 1978) and whetber the effect requires multiple &sting as opposed to merely extended retrieval durations (Erdelyi and Becker, ! 974; Roediger and Thorpe, 1978 : Shapiro and Erdelyi, 1974). A sign&ant feature of this program of research is that unlike pas’cwork on incremental recall effects, including the once active literatures on remiand hypnoti,: hypermnesia (cf., Erdelyi and Kleinbard, 1978), report criterion shifts over successive recall trials have been con trolled a forced recall procedure requiring subjects to generate the same number of responses in each trial., guessing if necessary. Thus, the obtained improvements in recall performance represent genuine increments in accessible memory, not merely increases in reporting (cf., Erdelyi and Becker, 1974; Zrdelyi and Kleinbard, 1978). With hypermnesic recall established as a bona fide memory effect, the question arises whether hypermnesia is restricted merely to recall, or whether it constitutes a more generalized memory phenomenon including, for example, remgnition. In the present study we sought to address this question by attempting to induce hypermnesia in recognition memory. What may on fmt glance appear to be a straightforward experimental problem-one requiring merely the transposition of the traditional multitrial recall design to multi-trial recognition -turns out in practice to pose sevem methodological as well as conceptual difficulties. To the extent, for exaqlple, that hypermnesia is restricted primarily to pictorial or imagistic stinMi, the notoriously high capacity of picture ree*+@tion memory ckerson, 1965; Shepard, 1967; Standing, Conezio and Haber, 1970) would tend to pose insuperable ceiling effect problems for this type of mvtigation. In an unpublished study by the senior author involving 120 piaorial stimuli of the type used in recall hypermnesia expetiments, subjects evidenced MuaLly perfect recognition memory (99% hit rates for 1% false alarm rates). Such near~erfeti initial performance would obviously preclude any detectable increments in subsequent recognition trials. The theoretic&y unproeg expedient of dampening picture recognition memory through either very r’ap?dpresentation rates or through the use of hard-todiscrimmate distractom fade, y in pilot studies, to yield any recognition increments in repeated rec0gnition tests.

Recognition Hypemnesia

25

It was clear to us that, in line with Erdelyi and Becker’s (1974) retrieval search explanation of hypermnesia, the key tc inducing recognition hypermnesia was a recognition task in which retrieval search played a non-trivial and active role. For this purpose we sought to move away from the traditional laboratory procedure of testing recognition memory with items identical to those in the original stimulus set, a task *whichin any case has questionable ecological validity, since few events in the real world are ever repeated identically, requiring us to recognize past occurrences from variants or from merely components of those events. We wished to adopt d class of highly configured stimulus materials such that the presentation of a component element would tend to yield, with some retrieval effort, the accessing of the whole stimulus and, therefon,, the recognition of the element as a constituent of the whole stimulus. In this type of recognition task, therefore, subjects would be asked to recognize elemelzts of presented stimuli rather than the stimuli themselves. A further requirement, in view of the differential effects obtained in the past with pictorial and verbal materials, was that configured stimuli be made up of both verbal as well as pictorial components, so that recognition memory might Le evaluated for both classes of materials. A type of stimulus that impressed us as being ideally suited for these purposes was cartoons with captions. Such stimuli meet not only the requirement that they be made up of both pictorial and verbal components, but also that these components be highly configured (i.e., for the stimulus to be a proper funny cartoon, the correct verbal caption must accompany the correct picture). A further advantage is that such stimuli can be easily disconfigured by the incorrect pairing of captions and pictures, allowing the construction of non-configured stimuli from the same population of elements (e.g., Figure 1). We were thus able to select a large number of configured (funny) cartoons and to create from them, by scrambling pictures and captions, a parallel nonconfigured (non-funny) set. It became possible, therefore, to assess multitrial recognition memory for both pictures and words from both configured (funny) and non-configured (non-funny) stimuli. Method

Design The experiment entailed a 2 X 2 X 3 factorial design with correlated measures on the last factor: stimulus (pictures versus captions) X configurations (configured or funny versus non-configured or non-funny) X recognition test (B,, Bz, Bs).

26

Midtew

H.

Emlelyi und Judy 8. Steh

Examples of conjigwedcartoons (left panel) and nonconfigured counteqmrts (r&h t punelj.

“Ye Gods, dbdufhh .!I”

‘Your

sfmfegic

fdhre,

wifhhuwul

wus

U

Genefu/!/"

The subjects, volunteers from introductory psychology classes at Brooklyn College, were tested in small groups of from 2 to 4 subjects each. To equalize ns for each of the four independent conditions (Pictures/Funny, Pictures/ on-Funny, Caption$Gunny, Captions/!Von-Funny), subjects were randomly dropped from any tirt-dtmentgroup exceeding n = 16. Thus, the analyses are based on four groups of 16 subjects each, for a total N of 64.

A total of 182 cartoons ?tith captions were selected fern a variety of m?lgazincs (e.g*, PErryboy, 7%~ New Yo&r) and joke books (e.g., Dennis t&v

Recognition Hypetmnesia

Menace, The

Thing

Between

27

Man and Women) and made into slides. Another

182 (non-funny) slides were made from scrambled versions of the original cartoons in which the 182 captions and 182 pictures were randomly recombined in inappropriate pairs. From each set of 182 slides, a subset of 140 funny, and 140 non-funny slides were selected to serve, respectively, as the funny (configured) and non-funny (non-configured) stimulus lists. The remaining subsets of 42 funny and non-funny slides served as a pool for distracters (foils) in later recognition tests. kecognition tests were based on a subset of 20 of the original 140 cartoons (funny or non-funny) plus 20 distractors chosen from the corresponding set of 42 non-stimulus slides. The same 20 stimulus and distractor items were used in each successive recognition trial. The 140 stimulus slides (funny or non-funny) were ordered in a projector carousel alphabetically according to the last word of each cartoon’s caption. All subjects were shown a complete set of 140 stimuli (funny or non-funny) at the rate of 15 set per slide. Procedure

At the beginning of the experiment all subjects were told that the study was concerned with their ability to remember a very large number of cartoans. They were to pay close attention to each slide as they would be tested on the materials later. Following the full sequence of 140 slides (which took some 35 min to present), the room lights were turned on and subjects were given instructions for their fust recognition test, RI. Each subject was &en a recognition booklet containing 40 cartoon pictures or 40 cartoon captions, numbered l-40. The subjects were told that half of the pictures/captions had actually appeared in the stimulus list while the other half had not. The subjects were requested to indicate on a provided recognition form whether they recognized (Yes/No) each picture/caption and, further, to indicate their recognition confidence (Positively No, Probably No, Probably Yes, Positively Yes). In essence, then, subjects produced a 4-point rating of their confidence of recognition of each of the 40 test items. The subjects were instructed to ate each test item in order, and were not allowed to review previous pages in the test booklet, nor to change their ratings. Subjects were allowed 10 min to rate the 40 test items, though a 2 min grace period was allowed if needed. At this point the test booklets and the rating protocols were collected and subjects were i.nstructeiLto sit quietly for 5 min and to use this time to think back to the stimulus list. This think period has been a standard feature of the typical recall hypermnesia procedure (and has been shown to augment recall hypermnesia (Erdelyi and Becker, 1974).

28

MatthewH. Erdelyi and Judy B. Stein

At the end of the 5 min think inl,erval, a second recognition test, Rz, followed, identical to the first (comprising the same stimulusdistractor set), followed by a second 5 min think interval, followed finally by a last, again identical, recognition test, RO.

Results and discussion The recogniticq ratings of subjects were transformed into hit and false alarm rates from wnich ROC curves were generated (Green and Swets, 1966). Figure 2 presents the average (group) ROC curves for the Pictures/Funny (P/F) and the Pictures/Non-Funny (P/NF) groups for each of the three MCcessive recognition tests. Figure 3 presents the corresponding ROC curves for the Captions/Funny (C/F) and the Captions/Non-Funny (CJNF) qroups. The results of the experiment are readily apparent from visu,! stspection of the four ROC triads. A systematic recognition memory increjnent was obtained in the P/F group but in none of the other groups. Thus, recognition hypermnesla was successfully induced for pictures from configured stimuli, but not for pictures from non-configured stimuli, nor for words from either configured or non-configured stimuli. An incidental but nevertheless striking feature of these data is the overall superiority of recognition memory for stimulus elements of configured over non-configured stimuli, regardless of whether these elements were pictures or words. Clearly, picture recognition memory need not be superior to verbal recognition memory; indeed, as comparison of the C/F curves (Figure 3) with the P/NF curves (Figure 2) will immediately indicate, verbal recognition memory can easily be superior to pictorial recognition memory. The original context of the tested materials, as shown in this study, is the critical determinant of salbsequent recognition levels. Statistical analyses bear out these various observations. An analysis of variance was performed on d’ indices of recognition memory cmxtractedfrom each recognition rating protocol. The d’ index used was the avU?ageof the (9’ values associated with each of the three criterion points (whiGh should, theoretically, be equal) generated by the $-point rating system. The d’ values for each criterion point lvere taken from Table I of Swets (1964). A 2 (stimulus) X 2 (configuration) X 3 (recognition test) analysis of variance with repeated measures on thy last factor (Winer, 1962) ,pielded a significant’ configuration effect (funny versus non-funny), F (1, 60) = 39.61, M’e = 1.99; a non+&nificant stimulus effect (p&u-es versus words), F (1, 60) = ‘The rejection region for all statistical tests isp < 0.61.

Recognition Hypermnesia

Figure 2.

29

The course of recognition memoly over three successiverecognition tests(RI, Rz, Rs) forpictures from configured (firnny) and non-configured (non-Jirnny) cartoons. lOO%PICTURES

/FUNNY

90% ;;; 3 ; Ii G 0: tz

80%TURESINON-FUNNY

70%-

60%-

I

Figure 3.

I

10%

20%

FALSE

ALARM

30%

40%

50%

RATE : P (Y/N)

The course of recognition memory over three successiverecognition tests(RI, R,, Rs jlbrcap . *ws from configured (funny) and non-configured (non-funny) cartoons.

CfiPTIONS/

FUNNY

90 %

(JO%

PTIONSINON-FUNNY

cl

60%

l

R3

0

R2 RI

0

FALSE

ALARM

RATE:

P(Y/N)

30

Matthew H. ErdeZyiand Judy B. Stein

1.13, MSe = 1.99; and a non-significant recognition test effect, F (2, 120) = 1.9 I, MSe = 0.102. The critical configuration X recognition test interaction was the only s&ficant interaction effect, F (2, 120) = 7.83, MSe= 0.102. A linear trend analysis revealed the recogrrition performance of the P/F group to be significantly incremental, F (1, 120) = 12.36, MSe= 0.102. The corresponding analysis for the C/F group revealed no significant increments, F < 1. AJI aIte.-nate analysis of variance was performed on d’ values based on subjects’ dichotomous Yes Versus No recognition responses (corresponding to the middle criterion point in the four-way rating scale). The analysis yielded the same paitern of results: a significant configuration effect, F (1, 60) = 3 1.42, MSe = 2.5 1; a non-significant stimulus effect, F (1, 60) < 1, MSe = 2.5 1; a non-significant recognition test effect, F (2, 120) = 1.66; MSe = 0-l 18; a s’ignificant configuration-recognition test interaction, F (2, 120) = 6.10, MSe = 0.118; and a significant linear trend in recognition performance overtesttrialsfortheP/Fgrouponly,F(l, 120)=8.97;MSe=O.118. The basic finding of this study, then, is that recognition hypermnesia may be induced over time with repeated testing. For recognition hypermnesia to be obtained, however, retrieval search must be a non-trivial component of the recognition task (as indicated by the failure of either the picture or word eli;ments of non-configured stimuli to increase over tests). Further, consistent with past findings of picture but not word recall hypermnesia (Erdelyi and Becker, 1974; Erdelyi and KleLlbard, 1978), recognition hypermnesia appears likewise to be restricted to pictorial stimuli (as indicated by the failure of captions from funny cartoons, in contrast with pictures from funny cartoons, to become more recognizable over time). A theoretical accounting of the present recognition hypermnesia results can be protided through an elaboration of Erdelyi and Becker’s (1974) retrieval-recognition explanation of recall hypermnesia. Recall hypermnesia, according to Erdelyi and Becker, occurs because each new recall trial as well as intervening think interval provides the subject with additional ti,me for the wtrieval-recognition o: orations that determine rc.:all performance. Specifically, the theoretical framework is a two-stage model of recall (e.g., James, 1890; Kintsch, 1970) in which the subject is assumed (1) to retrieve potential candidates for recall which are then (2) recognized or not recog nized as belonging to the stimulus set, resultin, in the overt recall or nonrecall of each considered candidate. The difference in net picture and word recall perfcrmance over time is explained by he recognition component of the recall process. Picture recognition under usual circumstances is exceptionally accurate (Nickerson, 1965; Shepard, 1967; Standing, Conezio, and Haber, 1970) but word recognition is relatively modest. Consequently, in the case of pictures, the retrieval of a correct candidate for recall almost inevi-

Recognition Hypermnesia

31

tably results in recognition, and hence in correct recall. Thus, increasing the time for retrieval-recognition operations inc (ritably increases recall of pictures. With word candidates, bos-lever, the recognition process is more chancy, with correct candidatea oeing frequently not recognized (resulting in 6‘misses”) and incorrect c‘ndidates being often recognized (resulting in “false alarms”). Thus, the new recalls resulting from further processing time are often incorrect (false alf ,ms) and those that are correct (hits) are insufficient to overcome underlyin; forgetting trends. Recognition hypermnesia can be explained by assuming that recognition performance, like recall, is also based on a two-stage process in which candidate items are first retrieved and then either recognized or not recognized covertly, resulting in cor3zspoinding positive or negative overt recognition responses (see, for example, Glass, Holyoak, and Santa, 1979; Mandler, 1972). Opportunities for further retrieval-recognition operations (additional recognition trials, intervening think intervals) should produce, as in recall, increases in performance in those recognition tasks where retrieval operations are not trivial, i.e., where retrieval of candidates is not essentially automatic but involves time-consuming search (otherwise, asymptotic performance would be reached on the initial trial with no room for improvement). Configured stimuli, because of the high association between their components, satisfy this requirement: unlike non-configured stimuli, a component of a configured stimulus, say a picture component, Pi, of a funny cartoon, suggests the retrieval space, Si, of its complementary component, the verbal caption, Ci, which, after some search, the subject may in fact retrieve, resulting in the reconfiguring of the complemen;tary components, P, and Ci, into the (highly recognizable) configured whole, [Pi + Ci] ; i.e., Pi + Si + Ci; Pi + Ci + [Pi +

Gil . Interestingly, the picture-woral difference observed in recognition hypermnesia needs to i)e explained through a picture-word difference in the retrieval component and not, as in the case of recall hypermnesia, in the recognition component of performanc-,p since recognition performance on both the captions and the pictures from the funny (configured) set are equivalently very high (see Figures 2 and 3). The retrieval difference, which very early became obvious to the experimenters, results from an asymmetry in the extent to which a non-recognized picture, Pi, in contrast with a nonrecognized caption, Ci, can generate the semantic domain (hence the ballpark retrieval space) of the complemen:ary element that would result in the reconfiguring of the whole (highly recognizable) cartoon, [Pi + Cl]. For example, a subject viewing the picture component, Pi, of the man-woman-baby cartoon (Figure 1, top left panel) m.ay, without immediately recognizing it as part of the old stimulus configuration, [Pi + Cl], surmise that it may have

31

MatthewH. Erdelyi and Judy B. Stein

something to do with an unwanted pregnancy, and from having narrowed down the retrieval space to this general theme, retrieve or reconstruct the pun&line, Ct, which now, in conjunction with Pi, is recognized (i.e., Pi is not recognized, But Pi + Si + Ci; Pi + Ci + [Pi -PCi] , and [Pi + Gil is recognized). The case of the verbal captions is quite different; if the caption, Ci (e.g., “‘Ye Gods, Abdullah!!), is not itself recognized or does not immediately lead to the corresponding picture, Pi, there is little chance that the subject will be able to reconstruct the pictorial theme from the caption, no matter how much additional time the subject is given (i.e., Ci * Si * Pipand thus, Ci * [Cj + Pi11. The determinant of the picture-word difference in hypermnesia is thus seer to lie with the recognition component of recall in recall hypermnesia, and wrtn the retriexl component of recognition in recognition hypermnesia. References ;‘r&lyi, M. H., and Becker, J. (1974) Hypermnesia for pictures: Incremental memory for pictures but agt words in multiple recall trials. Cog. Psych& 6. 159-171. Erdelyi. M. H., and KJeinbayd, J. (1978) Has Ebbinghaus decayed with time? The growth of recall (hypermnesia) over days. J. exper. Fsychol: Hum. Learn. Mem., 4, 275-289. Erdelyi. M., Buschke, H., and Finkelstein, S. (1?77) Hypermnesia for Socratic stinudi: The growth of recaB for an inter~~&~generated memory list abstracted from a series of riddles. Mem. Cog., 5, 283-286. Erdelyi, M. H., Finkelstein, S., HerreB, N., MiBer, B., and Thomas, J. (1976) Coding modality versus input modality in hypermnesia: Is a rose a rose a rose? Cog., 4,311-319. Glass, A. L., Holyoak, K. i., :md Santa, J. L. (1979) Cognition, Reading, Massachusetts, AddisonWesley. Green, 1). M., and Swets, J. A. (1966) Srj~naldetection theory and psychophysics. New York, Wiley. Madigan, S. (1976) Reminiscence ,and item recovery in free recall. Mem. Cog., 4.233-236. Mandler, G. (1972) Grganization and recognition. In E. Tulving and W. Donaldson (Eds.), &ganEation and memory. New Yark, Academic Press. Nickerson, R. S. (1965) Short-term memory for complex meaningful configurations: A demonstration of capacity. tin. J. Psychof 19.155-160. Roediger, H. L., and Thorpe, A. L. (1978) The role of recall time in producing hypermnesia. Mem. Cog., 6,296-305. Shapiro, S. R., and Erdelyi, M. H. (1974) Hypermnesia for ptitures but not words. J. exper. Psychol., Z03,1218-1219.

Shepard, R. N. (1967) Recognition memory for words, sentences, and pictures. J. verb. Learn. verb. Behav., 6.156-163. shiffiin, X. M. (1970) Memory search. In D. A. Norman (Ed.), Models of human memory. New York,

Academic&as. Standin& L., Conezio, J.. and Haber, R. N. (1970) Perception and memory for pictures: Single-trial karn@ of 2560 vi&l stimuB. Z=$ychoLScL, Z9,73-74 Swets, 3. A. (Ed.), (1964) j’ignclldetection theory and recognition by human observers. New York,

Wiley.

Reco@tion

Hypeimnesia

33

Winer, B. J. (1962) S ?ztistical principlesin experimentaldesign. New York, McGraw-Hill. Yarmey, D. (1976) Hypermnesia for pictures but not for concrete or abstract words. BuZ. psychol. Sot., 8, 115-117.

Rtkmt! Dans cette ktude on a cherche 1 voir si I’hypermn&ie de reconnassance (le desoubli) pouvait dtre obtenne en testant de faGon r&pc%tive la mimoire de reconnaissance. On a utilisk quatre types de stim&’ les images ou les mots (bulles) de bandes dessiraees avec correspondanc: de l’image et de la bulle { locc d&es) ou sans correspondance (pas dri3es). Apres prksentation d’une s6rie importante de bandes drs&es avec bulles, on a administrk successivement trois teqts de reconnaissance comprenant les mbmes stimuli et distracteurs. Entre les tests de reconnassance des intervalles de 5 minutes de reflexion silencieux &aient intercal&. On a obtenu une hypermn&ie de reconnaissance mesurGe par des estimations croissantes d’ de memoire de reconnaissance accessible, pour les images de bandes dessinees avec correspondances images-bulles (dr8les) et non pour les autres groupes.

Cognition, 9 (1981) 35-58 @Elsevier Sequoia S.A., La._sanne- Printed in the Netherlands

On the adequacy of prototype theory as a thmry of concepts* DANIEL N. OSHERSON Massachusetts lnstitu te of Terhnology

EDWARD E. SiUllTH Stanford

University

and Bolt Beranek and

Newman, trlc.

Abstract Prototype theory construes membership in a concept’s extension as graded, detemzined by similarity to the concept’s ‘3e.W exemplar (or by some other measure of central tendency). The present paper is concerned with the compatibi1it.J of this view of concept membership with two criteria of adequacy for theories of concepts. The first criterion Conrorns ;he relationship between complex concepts and their conceptual con;,&:CSrr The second concerns the truth conditions for thoughts correspo,ldi.T? lo simple inclusions.

A novel and ambitious theory has emerged from the last ten ye:rs of psychological research into the concepts that underlie ‘kind” terms like “animal”, “tree”, “tool”, and “clothing”. The distinguishing doctrine of the new theory is that entities fall neither sharply in nor sharply out of a concept’s extension. Rather, an object instances a concept only to the extent that it is similar to the prototype of the concept; the boundary between membership and nonmembership in a concept’s extension is thus fuzzy. We’ll call the new theory (really, class of theories) prototype theory. In this paper we consider two aspects of concepts relevant to choosing between pretcltype theory and its more traditional rivals. One concerns conceptual combination, that is, the process whereby relatively complex concepts are *‘I’%is research was partly supported by U.S. Public Health Service Grant MH-19705. We thank Ned Block, Susan Carey, D&vidIsrael, Louis Narens,Gary Olson, Molly Potter, Lance Rips, WilliamSalter, and Ken Wexler fcr helpful iliscussion of the ideas in rhis paper. Reprint requests should be sent to EdwardE. Smith, Bolt Beranekand Newman, Inc., 50 Moulton Street, Cambridge,MA 02138.

36

D. N. Oshelsonand E. E. Smi2h

forged out of relatively simple ones. The other deals with truth conditions for thoughts, that is, the circumstances under which a thought corresponding to a declarative proposition is true. For both aspects, we argue that the new theory of concepts fares worse than the old. The organization of this paper is as follows. We first present one version of prototype theory. We then show how it might be extended to account for conceptual combination by means of principles derived from ,&zy-ret theory (e.g., Z&h, 1965). This extension is demonstrated to be frautrht with difficulties. We then move on to the issue of truth conditions for thoughts, again using fuzzyset theory as a means of implementing the prototype approach, ar?, again iemonstrating that this implementauon won’t work. In a final seLlion, we establish that our analysis holds for virtually any version of prototype theory, and consider ways of reconciling previous evidence for this theory 1.ith the wisdom of the older kind of theory of concepts. Section 1. Prototype theory

The versron of prototype theory expounded i:i this section appears implicitly and explicitly in several well known papers (e.g., Posner & Keele, 1968; Rosch, Simpson and Miller, 1976), and it is the simplest account we know. However, other work within the prototype theory tradition (e.g., Reed, 1972) directly contradicts it, and there are numerous alternative formulations of the theory that differ from our version in various details. Nevertheless, the version about to be specified is “prototypical” of prototype models in that it captures key ideas of prototype thee-y more successfully than rival versions, and it is only these key ideas that are crucial for our analysis, as we delnonstrate later (see section 4.2). 1. _r I Formal characterization I According to the present version of prototype theory, many concepts can be identified with (mental) representations of quadruples like (*), (*) (A, d, P, c) where A is a set of readily envisionable objects (real or imagined) called a conceptual domain; d is a function from A x A into the positive real numbers, called a distance metric;

On the adequacy of prototype theov

37

p is a member of A, called the concept’s pvofotype; and c is a function frolm A into [0, 11, called the concept’s characteristic function; and such that the folio wing two conditions hold: (1.1)

(A, d> is a metric space, i.e., 0fx-Wfy-U

(la) d(x,y)=Oiffx=y (lb) d(x, y) = ‘;l(y, x) (1.2)

( 1c) d(x, Y) + W, 2) (VxEA)(VyEA) d(x,

2 dk

2)

PI g d(y, PI + C(Y) d c(x).

The second condition requires that the closer an object is to its prototype, the more characteristic it is of the concept. We can illustrate with the concept bird. According to prototype theory it is ic’sntical to a mental representation of the quadruple (B, dbird9 Pbird , Cbird ), where B is the set of all readily envisionable birds (including robins, sparrows, and penguins);’ db+d is a function on pairs of such birds into real numbers (aSSi@iIX smaller numbers to pairs of similar objects, e.g., robin-sparrow, than to pairs of dissimilar ones, e.g., robin-penguin, thereby reflecting relative psychological similarity among elements of B); pbtid is some particular bird in B (usually taken ?o !.e the bird that has the average value on each dimension of the underlying metric space); and ci.,W is a function assigning numbers in [O, 11 to members of B in such a way that closeness to the prototype reflects greater “birdiness.” When construed in this way, the concept bird appears to meet conditions (1.1) and (1.2). And this construal is consistent with a number of empirical findings. In particular, consider the case where subjects give similarity ratings for pairs including either two instances of bird (e.g., “robin-sparrow”) or one instance and the term “bird” itself (e.g., “robin-bird”), and then these &ic space. Assuming that ratings are converted into a multidimensional ‘If you come across an ordinary robin that is new to your experience, is the theory committed to a change in your bird concept, in light of the new element added to B? No, the new robin was already envisionable, so it was in B to begin with. On the other hand, if you happen upon a very queer bird {e.g., one with tusks), not so readily envisioned, your concept may well change by expansion of B.

38

D. N. Oshersonand E. E. S with

the point in the space for the term “bird” corresponds to tL2 prototype, the following fmdings emerge: (a) The value of the prototype on each dimension is roughly equal to the average value of all scaled instances on this dimension (Rosch t!kMen&, $975). This provides some justification for identifying the prototype with an average, though it is pot really critical for the present version of prototype theory s (b) In general, the less the metric distance between any bird and the prototype, the more characteristic or prototypical of the concept that bird is judged to be (e.g., Rips, Shoben & Smith, 1973). Since different groups of subjects make the initial pair&e similarity ratings and prototypicality judgments, these results provide good evidence for condition (1 .Z) above, i.e., for the psychological reality of a function mapping distance from a prototype into a characteristic function ‘4at reflects “birdiness”. *z) In general, the more characteristic or prototypical a particular bird: (i) the more efficiently it can be categorized, efficiency being measured by both accuracy and speed (e.g., Rips et al., 1973); (ii) the earlier it will be output when subjects list instances of the concept (e.g., Rosch et al., 1976); and (iii) the earlier a child will learn that it is an instance of the concept (e.g., Rosch, 1973). Thus, the value of the characteristic function for an object can be used to predict aspects of real-time processing and conceptual development. 6.2 Some gap 6s ?he theory While the above fmdine provide some support for one version of prototype theory, there Cirecurrently two substantial gaps in this or any other extant vetion of the theory. First, not all natural concepts succumb to the kind of construal we have given for bird. Prototype theory, LS thus far developed, is bzt suited to “kind” notions (like &g, tree, and animal), to “artifact” notions (like tool and clothing), and to simple descriptive notions like triangular arm red What remains outside the theory’s purview are intentional or otherwise intricate concepts like belieA desire, and justice, as well as the meanings of prepositions, sentence connectives, and a host of other ideas. it is an open question whether or not the theory can be extended to cover these cases, and in this paper we shall say no more about them. The second gap is the inverse of the fast ; namely, many nonnatural concepts do succumb to the kind of construal we habe given for bird. Accord?g to the present version of prototype theory, each concept is idenllcal to some quadruple, with distinct concepts being identical to distinct

On the adequacy of prototype theory

39

quadruples (i.e., quadruples differing in at least one component).2 But many such quadruples will have bizarre domains (and perhaps other defects :,s well). It is possible, for example, to use a single metric space to represerit objects as diverse as dogs, chairs and toothpaste (thereby satisfying conditiot (1 .l), to define a prototype in this space, and to then map distances from this prototype into values of a characteristic function so that small distances go with larger characteristic values (thereby satisfying condition (1.2)). This state of affairs reveals an important incompleteness in the theory. We need to know more about the entities A, d, p, and c than is revealed by conditions (1.1) and (1.2), so that we can limit the theoretically possible concepts to those that are psychologically natural. Again it is an open question whether such an extension of the theory can be made, and again we shall say no more abet? t the issue here. Section 2. Prototype theory and conceptual combination

Despite the gaps ju st mentioned, it is useful to evaluate prototype theory by means of facts about conceptual combination, where the concepts are natural ones and restricted to kind terms. Though there has been some experimental work on this issi;:: (e.g., Hersch & Caramazza, 1976; Oden, 1977), tls present analysis provides a more general outlook. In what follows, we first briefly describe the issue of conceptual combination, &n relate tixns issue 20 legitimate demands tIpon prototype theory, next spell out the principles of fuzzy-set theory oti which prototype theorists rely in order to deal with conceptual combinaticn, and lastly show that this enterprise eventuates in a snarl of contradictions. 2. I The issue One or more concepts combine to form anoth::r whenever the latter has the former as constituents. Grammatical constituency can often sexve as a guide to conceptual constituency. Thus, the words “red” and “table” are constituents of the grammatical structure “red table,” and in parallel fashion the concepts Ired and table are constituents of the conceptual structure red table. The same parAle holds for many other conceptual combinations, e.g.,

‘More precisely, dlffment concepts are to be ideatiied with mentd representations of different differentmental representatonsof the same quadruple to be variants of the same concept). In what follows, we will occasionally suppress thin mental-representationquallller.

quadruples (we% count

40

D. iV. Oshemonnnd 17.E. Smith

square window and tasty onion. 3 There are, of course, cases where the paral-

lelism breaks down. Thus, the concept dark horse (as in political contests) does not have dark and horse as conceptual constituents. Such idioms notwithstanding, it seems safe in what follows to frequently rely on grammatical structure as a guide to conceptual structure. The phenomena surrounding conceptual combi,qation can be used to evaluate prototype theory in the following way. Let’s say that this theory is compatible with conceptual combination if principles can be supplied that correcc;y predict the relation between complex concepts an+: their constituents, when concepts are construed as specified by the theory. Or at least, many such conceptual combinations accord with the theory’s principles and few violate them. 2.2 Criteria of adequacy regarding conceptual combination Suppose a given concept C has concepts C1 and Ca as constituents. For prototype theory to be compatible with this case of conceptual combination, principles must be available to specify the quadruple associated with C on the basis of th,? ,duadruples associated with C1 and C,P Actually, these principles ought to specify the (mental) representation or the former quadruple on the basis of the (mental) representations of the latter quadruples; but in the present em of psychological understandmg, we shall settle for a specification, not of the representation (i.e., not c ‘C itself), but of the associated quadruple (i.e., of what C represents). The desired principles, then, will characterize C’s domain, distance metric, prototype, and characteristic function on the basis of the entities from the quadruples for CI and Cz. However, we shall restrict attention to the problem of specifying C’s characteristic function on the basis of those of C1 and CZ . There are two reasons to so narrow the present inquiry. First. a solurion to the character-i& function problem is necessary for a gene&Asolution to the problem of conceptual combination, and it may be close to sufficient as

?It is unclear whether the parallelholds for grammaticalstructures l&e “very happy”. It may be that very kpp~~ has only one conceptual constituent, namely, Iwrpp~.That is, the word ‘very” may not ¬e a concept in very Iazppy,since it may not correspond to a constituent in the language of thought, but instead be representedsyncategorematic4ly. Simiku remarksapply to UnhuppY.Well extend the ordinary usage of “comtination” to atlow a sin@I collcept to be “combined” into another concept by the use of such devices. 4 Additionally, it must be true that prototype theory successfblly construes C, Cl, and (2% as mental represcntatiok4of quadruplesof form (*) in the firpltplace. Well assume 80 in what follows.

On the adequacy of protorype theory

41

well? Second, the only explicit account of conceptual combination within the prototype theory tradition (given below) restricts its attention to the characteristic functions of C, C1 , and C1. We may now investigate the compatibility of prototype theory with the facts of conceptual combination, in particular, with the relationship between the characteristic functions of complex concepts and the characteristic functions of their conceptual constituents. Combinatorial principles germane to this problem have been supplied by Zadeh (e.g., 1965) under the name of fuzzy-set theory, and prototype theorists (e.g., Oden, 1977; Rosch and Mervis, 1975) cite fuzzy-set theory in this connection. It is possible that principles other than those provided by Zadeh can better serve prototype theory in accounting for conceptual combination, but no suitable alternative has yet been suggested (so far as we know).6 Since, in addition, fuzzy-set theory is a natural complement to prototype theory, and the former is an appealing theory in its own right, we shall evaluate prototype theory esclusively in the context of fuzzy-set theory. If this ensemble is at variance with the facts of conceptual combination, there is some reason to doubt the compatibility of prototype theory with these facts. 2.3 Fuzzy-set theory The principles of fuzzy set theory are straightforward generalizations of elementary principles of standard set theory.’ We start by enumerating the relevant principles of standard set theory. Let D be the domain of discourse, and A a subset of D. In standard set theory, the characteristic finction for A is defined to be that unique function (2.1) CA:D + (0, I} such that (2.2)

(V x E D)cA(x) = ;iff;;; I

.

s For: (a) C’s domain is likely just the union of those for Cl and C2, (b) C’s distancemetric can be highly constrained by its characteristicfunction if condition (1.2) is strengthened in any of several empirically plausible ways (we omit the details), and (c) c’s prototype can be taken as the “average” memberof C’s domain according to its distance metric. ’ In Sections 3.5 and 4.3 we discussalternaiivesto Zadeh’s(1965) proposals. ‘Zadeh (1%5) develops fuzzy-set theory beyond the limits imposed here, e.g., to deal with the composition of relations. However, we will present enough of the theory to assessits empiricaladequacy.

42

D. N. Osherson and E. E. Smith

Set membership is thus strictly binary; there are no liminal cases of objects falling neither precisely in nor precisely out of a set. The characteristic functions for intersections, unions, and complements of sets can be standardly defined as follows (where A and B are arbitrary subsets of D). (2.3) (2.4) (2.5)

(VXED) (vx~D) (VXED)

c An B(x) = min(c* (x), cg (x)) (Intersection) c A U B(x) = max(cA(xh (union) 63 (x)) (Compiement) Cnd(X)= 1 -CA(X)

The characteristic functions for the empty and universal sets,@ and D, meet conditions (2.6) and (2.7), respectively. (2.6) (2.7)

(V x E D)CS (x) = 0 (Empty set) (V x E D)cn (x) = 1 (Universal set)

Thus, in the domain of animals, suppose that cdW(Rover) = 1 and cfanale (Rover) = 0. Then: cd@f,&(Rover)

= minicd,(Rover), cfati,(Rover)) I*‘” = min(1, 0) = 0; cdaUfenale(Rover) = max(c&,(Rover), cfmale(Rover)) =max(l,O)= 1; = 1 - cd&Rover) = 1 - 1 = 0; c,,tig(Rover) c,demde(Rover) = I - cfmale(Rover) = 1 - 0 = 1. Fuzzy set theory results from expanding the range of characteristic functions from {O, 1) to [O, 1 I , i.e., from the set whose only elements are 0 and 1 to the set of all real numbers between 0 and 1, inclusive. Set rnelmbership thus becomes continuously graded, an element, e, belonging to A to the extent that c,(e) approximates 1. Instead 01 (2.1) and (2.2), we now have (2.8) (2.9)

CA: D+ [0, 1] The larger CA(x), the more x belongs to A; the smaller c,(x), the less x belongs to A; 1 and 0 are limiting cases (for all x e D).

The definitions given by (2.3~(2.7) remain unchanged, but the characteristic functions therein are now taken to be of the fuzzy-set type, (2.8), rather than of the standard type, (2.1). To illustrate, if cdW(Rover) = 0.85 and cfenat(Rover) = 0.10, then:

On theadequacyofprototypetheory

c&g

43

nfemdRoveO= mWdJRover), cfemate(Rover))

CdogUBemale(Rover)

c,,h,(Rover) c,demale(Rover)

= min(0.85,O.lO) = 0.10; = max&,JRover), cfemde(Rover)) = max(0.85,O.lO) = 0.85; = 1 - c&,,(Rover) = 1 - 0.85 = 0.15; = 1 - cfeemale(Rover)= 1 - 0.10 = 0.90.

2.4 Is prototype theory compatible with conceptual combirzation? We are ready to demonstrate that prototype theory in conjunction with fuzzy-set theory contradicts strong intuitions we have about cencepts. Three problem areas are considered: conjunctive concepts, logically empty and logically universal sets, and disjunctive concepts.* 2.4.1 Conjunctive concepts Let the domain of discourse be the set, F, of all fruit, and consider the characteristic functions for the (fuzzy) concepts apple, striped, and striped apple. The concept stripe< apple stands in the same relation to striped and apple as does red house to red and house, square field to square and field, and so on. Striped apple thus qualifies as a conjunctive concept. Within fuzzy-set theory, conjunctive concepts are most naturally represented as fuzzy intersections, as in (2.3). At least, they clearly are not fuzzy unions, nor any yet more complex fuzzy Boolean function of their constituents. So, if there is to be any account of the conceptual combination striped apple within fuzzyset theory, it will likely rest upon (2.10). (2.10) (Vx E F) (c,~ -re(x) = mh-&&&x), c,,l,(x)) Equation (2.10) exhibits striped apple as a familiar kind of combination of the concepts striped and apple. Without (2.10) it is doubtful that fuzzy-set theory can secure the compatibility of prototype theory with this elementary case of conceptual combination. *A comment is in order about the kinds of problems that we will not consider. One kind involves counterexamples to fbaxy intersections that can be generated with concepts like connfe$eit, bogus. tmitutfon, etc. For example, a “good” counterfeit dollar, d, might be a “good” counterfeit, but it will be a “bad”’ dollar; hence, the characteristic function for the relevant fizzy intersection between countetj’kit and dollar will mistakenly declared to be a “bad” counterfeit dollar, since the intersection function yields the minimum value of the constituents. We wig pass over this kind of objection, however, in view of our earlieragreementto treat only simpleconcepts, among which counterfeit and the like cannot be found.

D. M. Oshemn and E. E. Smith

44

(b)

(cl

But (2.10) cannot be maintained. Let (a) in Figure 1 be a particular apple in F. There can be no doubt that it is psychologically less prototypical of an apple (whose prototype looks more like (b)) than of an apple-with-stripes; SO, :.2-11)

&biped

appie 00 > Gppk (a).

heqrsa?ity (2.11) asserts, simply, that (a) is a better illustration of the concept strii~ed-apple than it is of the concept apple (just as the reverse is true of (b)). 33ut(23 1) contradicts (2 .lO) since the latter implies

Tk

ht?Wd

a&W? (a) = mh

which in turn implies

h&&ed

(ah %Wle (a))

On the adequacyof prototypetheory

(2.12)

cstrlped apple

45

(a) Q capple (a).

And (2.11) and (2.12) are inconsistent. Given the above, it becomes clear that prototype theory conjoined to fuzzy-set theory will lead to a contradiction whenever an object is more prototypical of a conjunction than of its constituents. When phrased this way, numerous familiar conjunctions seem to provide counterexamples, e.g., a guppie is more prototypical of the conjunctive concept pet fish than it is of either pet or &kg We conclude that prototype theory cum f&y-set theory is not compatible with strong intuitions about conjunctive concepts. 2.4.2 Logically empty and logically ur~ivcrsad ccncep is The concept apple that is not an apple is logically empty since it can apply to nothing.‘* Its characteristic function within fuzzy-set theory should reflect this peculiar property; that is, its function should conform to conditiol? (2.6), which in this case can be written as: (2.13)

(Vx E FSc apple that id not an

apple(X)

= 0.

However, within fuzzy-set theory, the concept apple that is not an apple would be repre:;en:ed 1s the (fuzzy) intersection of apple and nonappZe yielding (by (2.3) and (2.5)): (2.14)

(V XE F’)Capple that is not an apple(X) = mh(Cappk (X), 1 - CappIe(XI)* Consider again apple (a) of Figure 1. Since (a) is a “better” apple than (c), but a “worse” one than (b), we have, by condition (1.2):

(2-l 5) cappie(c) < cappIe(a) < CappIe(b). (2.15) implies (2.16) (in the presence of (2.8). (2.16)

O< cappP,(a)< 1

gIn derivingour counterexample, we chose to work with an unfamiliarconjunction (stripedapple) rathtr than with a familiarone (e.g., per fi&) so as to make it less likely that the readerwould have representedthe conjunction as a single conceptual constituent. The latter could be the case for many familiarconjunctions (Bolinger, 1975; Potter & Faulconer, 1979), and for such cases representationIn terms of intersection is misleading. “Some might object on the grounds that people frequently use locutions of the form “-and not a -” to describec me object. For example, one might describetomatoes by saying: (i) They are both fruit and not fruit. But (i) seems to be a case in which grammaticalstructure is a misleadingguide to conceptual structure; (i) is probably idiomatic for asserting that tomatoes have some properties of fruit but lack others. isa_-and not a ---*’ that we can think of More generally, all locutions df the form “seem to be idiomatic.

46

D, N. OshersonandE. E. Smith

But if the appleness value of (a) is between 0 and 1, then so is one minus that value. Therefore both cagple

Cliwaitity

Gweatment(C)

>

Gnve*tment(~)

(i) ~wealth(B)

>

Cwealth (A)

(6)

>

Cwealth (C?.

Cwea~th (B)

(B)

Now if fuzzy-set theory is to represent the conceptual connections among liquidity, investment, and wl,alth, it would seem that its only option is to employ fuzzy union, as in (2.22). (2*22)

(Vx E D)cW~alth(X) = Cliquidity

U hvestm-nt

(Xl-

In light of (2.4), defining union, (2.23)

(V x E D)c,,titb 00 = max(CUquidty(xl,

Ci*ve&nent(X))-

But (2.19~(2.23) are inconsistent. To see this, focus on Cam*. According to (2.23), cwedti (B) is the larger of cliqUiddt,,(B) and chvestment(B). Suppose that cUqUidity (B) is the larger, then: (2.24) Now

than

cliquiaity

(W

= cwealth

(B).

focus on cwsolth(A). Since, by (2.23) again, cwealth(A) cannot be smaller cuquimy

(2.25)

(A),

c,mm(A)

2

Cu,uiaity(A)-

Combining this information with (2.19) yields: (2.26) And

cwarlth (2.26)

(A)

implies

> cUq&ity

(A)

> ctiquiciity (B)

= cweelth

(B).

48

D. N. Oshenonand E. E. Smith

Cweatth

(A) >

Cwealth

@It

co;,tradicting (2.2 1) (i). The only alternative is that Cbvesment(B)is equal to or larger than Cgguid&y (B), making the former the value of C,ede(B). But then, by b parallel argument to that just given, we have: which contradicts (2.21) (ii). Hence fuzzy-set theory does not properly represent the relation betv een wealth on the one hand, and liquidity and inves#me~t orI the other. We conclude that fuzzy-set theory renders prototype theory in::ompatible with strong intuitions about disjunctive conceptual combination. Section 3. Prototype theory and the truth conditions of thoughts

3. I The issue We take concepts to be the immediate constituents of thoughts. Given

this, it seems reasonable to ask a prototype theory of concepts for principles that informatively characterize the circumstances under which thoughts aE true, (when the thoughts correspond to declardtive propositions). In particular we are concerned wrth whether prototype theory, coupled with fuzzyset theory, can offer a cxnxt accourlt of simple quantificational thoughts like All A ‘s are B’s, Somt A ‘s are B’s, No A ‘s are B’s, and so forth. We can restrict attention to All A ‘s are W’s since our remarks will apply mutatis mutandz3 to other simple qsantificational thoughts. 3.2 Criteria of adequacy regarding t&h conditions

Let A and B be concepts compatible with the claims of prototype theory, and let T be the thought expressed by (3.1). (3.1)

All A’s are B’s_

On prototype theory, T consists, in part, of (mental) representations of the quadruples for A and B. A fully adequate version of the theory would supply principles that specify the truth conditions of T on the basis of these latter representations. Since so little is known about (such mental) representations, we shall require only that prototype theory specify the truth conditions of T on the basis of the quadruples themselves, i.e., on the basis of A and B’s domains, distance metrics, prototypes, and characteristic functions. And for reasons similar to those advanced in Section 2,2, we shall focus on the

On the adequacy of prototype theory

49

availability of principles that relate A and B’s characteristic functions to T’s truth conditions. In summary, then, we require of prototype theory a specification of what makes (3.1) true on the basis of the characteristic functions for A and B. Once again it is fuzzy-set theory that saves prototype theory from inexplicitness. Zadeh (1965) offers an explicit principle relating inclusions like (3.1) to the characteristic functions for A and B. Since Zadeh’s principle naturally extends the principles reviewed earlier, since his proposal has interest in its o In right, and since no alternative principle has been advanced to fill the prest& theoretical need (so far as we know), we feel justified in pinning the 1 lpes o? prototype theory to the fuzzy inclusion principle. If this ensemble is at variance with strong intuitions pertaining to the truth conditions of thoughts, there is reason to doubt the compatibility of prototype theory with those intuitions. 3.3 Fuzzy-set theory again Fuzzy inclusion is a straightforward generalization of standard inclusion. Standardly, the inclusion (3.1) is assigned the truth condition (3.2). (3.2)

(vx E D) (c*(x) Q cg(x))

To illustrate in the domain of animals, (3.3)

All females are dogs

is true just in case (3.4) is true. (3.4)

(Vx E animals) (cism&x) Q ci&)).

\~~‘,co;~T

,

)ou\false in light of a female nondog, a, for whom cfemale(a) =

The fuzz”v=mclusion principle results as before from generalizing the notion of a’characteristic function from’(2.1) and (2.2) to (2.8) and (2.,9) The fuzzy truth condition on (3.1) may still be stated as (3.2). Th,.:n, (3.3) remains false in fuzzy-set theory, in the light of a female rabbit, r, for whom chrnd,,(r) = 0.80 and cd&) = 0.20, contradicting (3.4). .3.4 Is prototype theory compatible with the truth conditions 2finclusion ? We now present a counterintuitive result to which prototype theory, m the context of the fuzzy inclusion principles, is committed.12 In this example, we use the domain, A, of animals, and the subsets associated with grizzly bear and inhabitant of North America. Consider the following inclusion. l2For footnote please see overleaf.

50

D. ik! Oshersonand E. 6 Smith

(3.5)

All gti&

!..JZJBan inhabitants of North America.

Within fuzzy-set theory, the truth condition of (3.5) is

(3.6)

(V XE A) (cglri~zly bear (x) d &habitant of North Amer~a 00). One thing we want from (3.6) is that it capture our iM.d:ions about what conditions insure 1:hefalsehood of the inclusion (3.5). P rt (3.6) fails to do this. For: according to (3.6), the existence of a squ’;rel on Mars is sufficient to falsify the proposition that all grizzly bears live in North America. The argument unfolds as follows. The characteristic function for grizzly bear is supposeo to represent the degree to which animals count as grizzly bears. Thus, polar bears will be assigned a numbqer close to 1, frogs ‘1number closer to 0, and earthworms a number yet closer to 0. Let Sam be a squirrel, and suppose that (3.7)

c wziY b&Sam) = p (where p > 0).

The characteristic function for inhubitant ofNorth America, if it is to accord with our intuition. must assign prcgressivc-ly smaller numbers to locations that are progressively further from the North American continent. Thus, it assigns Cuba a higher number than Scotland, which in turn is assigned a higher number than is Egypt. Now, is there a place that Sam could live so that cinhabitant would assign Sam a number smaller than p? If of North America so, we will contradict (3.6) and hence falsify (3.5). Surely, if Sam lived far enough from North America, the value of cinhabitant of North America (Sam) would be forced below p. l3 Perhaps a home in Egypt would suffice; if not, we can imagine he lives in Pakistan or Indonesia, or, if necessary, on Mars or “Another comment about kinds of problemsnot to be considered: It is possible to produce cmntelc examples to fuzzy inclusions by suitable choice of the domain of discourse. In a domain including both galaxiesand radishes,for example, the false inclusion All smallgalaxiesare small things

.

would be declaredtrue by fuzzy set theory because it is a theorem of the theory that for all subsets ,7B B,ofD,andallxED, We choose r,ot to rely on th& class of counterexample since the problem of relative ad(jectlveslike s??uJZZ is too severe a burden to impose on a theory with so few competitors. Consequently, in the examples to follow, we employ natural(in particular,intuitively homogeneous) domains. ‘junless cinaobitant of Noxth Amer k a asymptotes above p. But this is an arbitraryand counterintuitive restriction on the behavior of ciuhpbit& of North Am&.a; it would moreover, complicate the repzsentation of subjects* relative similarity judgments, since the psychological distance between objects far away from North America does not seem suMciently foreshortened compared to those nearby for such an elevated asymptote.

On the adequacy ofpmtotype theory

51

Pluto, in the vicinity of a nearby star, or within some distant galaxy. Let’s suppose that Mars is far enough. Then we have: (3.8)

%habiwt

Of

North America(Sam)

< Pv

Putting (3.7) and (3.8) together, we have: (3mg) c&zzly bear (Sam) > Cfnhabitant of North Amerlea(Sam)9 which contradicts (3.6). Since it is easy to generate an indefinite number of examples like this, it seems that fuzzy-set theory does not render prototype theory compatible with the truth conditions of inclusions.14 3.5 Partial falsification to the rescue ? One response to the difficulty just exhibited for fuzzy inclusion is to embrace fuzzy truth, that is, an (uncountable) infinity of truth-values between 0 and 1. On this view, the inclusion (3.5) is not outrightly falsified by Martian squirrels but only slightly weakened, some scheme being envisioned for computing the falsifying effects of counterexamples to truth conditions of form (3.2). To support this proposal it is pointed out that the discovery of Martian squirrels would indeed lower our confidence that all grizzly bears inhabit North America (if Martian squirrels, why not other Martian mammals?). So, psychologically, at least, (3.5) is partly falsified by (3.9), as the new approach predicts. Nene of this is well taken. We give thEe reasons for our distrust of a fuzzy-truth remedy to the problems associated with fuzzy inclusion. Note first that the new pr*-rposalis no mere adjustment of truth condition (3.2), found inadequate i,! the last subsection. To replace (3.2) with a “partial falsification” scheme so as to allow for truth values between truth and fatsity requires nontrivial trleory construction of a kind apparently not yet undertaken. Such a theory will need to include principles governing the degrees of falsification associated with (partial) counterexamples to inclusions. Reverting to our example, the theory must decide (a) how much addi-

t4An inleresting class of examples results from considering inclusions, A 5 B, such that A is a very atypical kind of B, e.g., All penguins are birds. All tomatoes are fruit. All bats are mammals etc. For, there may well be some quintessential penguin that is a better examplarof penguin than of bird, thereby falsifying the generalisation that all penguins are bir 3 (and similarly for the other castes).

52

D, IV. Oshersonand E. E. Smifh

tional falsification is engendered by the existence of n + 1 Martian squirrels instead of n (for all n), (b) whether some number of Martian squirrels falsify (3.5) as much as one European grizzly bear, (c) whether, if giraffes count as grizzly bears no more nor less than do squirrels, then a Martian squirrel and a Martian giraffe falsify (3.5) to the same extent as two Martian squirrels or two Martian giraffes, and so forth for many obvious questions. Until such a calculus cif partial falsification is offered that provides principled answers to questions like these, no explicit alternative to truth condition (3.2) is available to colrjoin to prototype theory. Qcond, and in the same vein, allowing propositions to admit of an inftity of possible truth values raises problems for the interpretation of sentential connectives and quantifiers. This issue is analogous to that for conceptual combinatiion, only now it is the combination of simple propositions into complex propositions that concerns us, instead of the combination of simple concepts into complex concepts. In particular, the familiar semantical apparatus of classical logic needs to be replaced since standard semantics is appropriate for interpreting its connectives and quantifiers only in the context of binarily valued propositions. To interpret, e.g., the conjunction of infmitarily-valued propositions, an appropriate infinite-valued Logic needs to be invoked. Many such logics have been devised; see Rescher (1969, pp. 3645) for a survey, as well as Zadeh’s (1975) own preliminary paper. The problem is that infinite valued logics generally violate strong intuitions about truth, validity, and consistency.15 At the least, their psychological suitability is not obvious. The advocate of fuzzy-truth, then, incurs the responsibility of selecting and defending an appropriate infinite-valued logic, and this promises to be a nontrivial undertaking. But until such a logic is presented, a serious alternative to fuzzy-set theory is not available to conjoin to prototype theory. Finally, we suspect that the partial falsification proposal results from mistaking degrees of belief for degrees of truth. We share the intuition that “To illustrate with an influential system, consider Eukasiewicz’sinfinite valued logic C-aleph. The intuitively valid sentence If John is happy, and if John is happy only if business is good, then business is good. if NM nontautologou s in .JUeph. (A tautology within R-aleph is any formula that assumesmwimal truth value under all truth assignmeutsto atomic subformulas.)Also, explicit contradictions (of the form (p and -p)) may differ in truth-value.Many other counterintuitlve results issue fromX-aleph The rivalinfinite valued lo&s in the literatureappearto yield aimibuparadoxes. None of these paradoxical results are features of classical (two valued) logic, although it is well known that classicallogic yields some counterintuitive results of its own (for discussion, see Kreuger and Osherson,forthcoming).

On the adequacy of prototype theory

53

the discovery of Martian squirrels should lower confidence in the truth of (3.5), but it is the confidence that is partially weakened, not the truth. Analogously, evidence migh? be adduced that alters our confidence in some as yet undecided mathem-+: ,-_,cal assertion; but the assertion, irrespective of our doubts and conjectunus, is either flatly true or flatly false. In other ‘words, the effect on our beliefs of the discovery of Martian squirrels is a problem relevant to inductive logic, not to the analysis of the truth conditions of inclusions. The “supporting intuition” of the fuzzy-truth advocate thus explained away, it is helpful to appeal to another intuition, namely, that it is logically possible both for all grizzly bdars, without exception, to dwell in North America ant’ for Mars to be teeming with squirrels; the irrelevance of Martian squirrels to the inclusion (3.5) - pace fuzzy-set theory and partial falsification schemes - is thereby revealed. We conclude that programmatic suggestions about partial falsification do not rescue prototype theory from its dependence on fuzzy-set theory; in particular, truth condition (3.2) is still the most natural and explicit interpretation of the inclusion relation within the prototype theory tradition; and (3.2), we saw, has unwelcome consequences. Section 4. Implication!f In this final section we consider three issues: (1) the “immunity” of standard set theory from the contradictions we derived, (2) the susceptibility of all versions of prototype theory to the derived contradictions, and (3) what can be salvaged from prototype theory in the light of these difficulties. 4. I Standard set theory and traditional theories of concep ts Standard set theory, with its binarily valued characteristic functions, is the natural complement of traditional theories of concepts -- at least insofar as those theories enforce a sharp boundary between membership and nonmembership in a concept’s extension. We now wish to point out that this amalgamation does not lead to the contradictions we derived from the amalgamation of fuzzy+et theory and our version of prototype theory. Hence, the former amalgamation is superior to the latter when it comes to accounting for conce#ual combination and the truth conditions of thoughts. The point is trivial to demonstrate. For, every contradiction we derived rests partly on the assumption that constituents of complex concepts can be assigned characteristic function values that fall between 0 and 1, i.e., this assumption is necessary (though not sufficient) for all our derivations. TO

54

D. N. OshersonandE. E. Smith

take the most obvious example, consider the concept we used in discussing logically empty sets, namely, apple that is not un apple. Recall that in fuzzyset theory the characteristic function of this concept meets (2.14): (2.14)

(Vx E F)

Capple that, in not an apple (XI

= mh

@apple (XL

1 -

cappie(x))

Recall further that the contradiction arose because there exists an apple whose appleness value is between 0 and 1, thereby insuring that one minus this value is also between 0 and 1: which in turn insures that the minimum is greater than 0. In standard set theory the characteristic function of apple that is not an apple must also meet (2.14), but since the appleness value of any apple must be either 8 or 1, for any x, one of CappIe( 1 - CappIe equaIs 0, thereby insuring that the minimum will be 0. A similar analysis in terms of standard set theory will remove the contradiction in the other cases we considered. 4.2 Exrension of the present results to other versions of prototype theory It was noted at the outset that there are versions of prototype theory alternative to the one we presented; indeed, some of these alternative versions enjcy greater empirical confirmatio:. than that we presented. There is, however, good reason to believe that the contradictions we derived apply to other versions of prototype theory as well. In the version we chose, concepts were identified with mental representations of quadruples like (*), (*) (A, Id,P, c), where A is a conceptual domain of envisionable objects, d is a metric function, p is a particular object in A, and c is a characteristic function. Most other versions of prototype theory differ from the above in their choices for the second and third coordinates of (*). Instead of a function that measures the psychological similarity between objects in terms of distance, many prototype theorists favor a function that measures this similarity in terms of number of common features (e.g., Rosch & Mervis, 1975; Smith, Shoben, and Rips, 1974), or in terms of a weighted contrast between common and distinctive features (Tversky, 1977). (The latter function has the virtue of not requiring that the similarity btitween two objects be symmetric.) And instead of the prototype being a particular envisionable object in the domain, some have assumed it is an abstraction from objects in the domain, e.g., a collection of features that occur most frequently in the domain (Smith et al., 1974). But such variations would not influence the contradict;?ps we derived. For as long as a version of (*) postulates that (a) there is some proximity

On the adequacy of pm tovpe theory

55

function that computes similarity between an object and a protot;:pical description, and (b) this computation results in graded similarity values that are mapped into a characteristic function such that high similarity goes with high characteristic values, then (c) the characteristic functions associated with the constituents of complex concepts can assume values between 0 and 1. And point (c), in conjunction with the formalism of fuzzy-set theory, is what produced all our contradictions. This can again be seen most easily by considenlag the logically empty concept: apple that is not an apple. Even if the prototypical apple is an abstraction, and even if similarity to a particular apple is computed by summing common features, it will still be the case that a particular apple can have an intermediate similarity value that results in a characteristic function value betwen 0 and 1. And this is all it takes to insure that the mirimum constituent of apple that is not an apple is greater than 0, which in turn insures that this logically empty concept is not associated with a chancteristic function that is identically zero. Similar remarks apply to other cases of conceptual combination we considerea, and to inclusion. As best we can tell, then, no current version of prototype theory is immune from the problems we have raised in this paper. 4.3 General implications.for pro totype theory One thing is clear. Amalgamation of any of a number of current versions of

prototype theory with Zadeh’s (1965) rendition of fuzzy-set theory will not handle strong intuitions about the way concepts combine to form complex concepts and propositions. This is an important failing because the ability to construct thoughts and complex concepts out of some basic stock of concepts seems to lie near the heart of human mentation. Where does this leave prototype theory? Three answers are worth considering. The first is that the theory is a sound rendition of concepts, but that it should beware of associating with Zadeh’s (1965) specific vers,ion of fuzzy-set theory since the fatter brings with it the kinds of problems we have discussed. This possibility receives some support iq Oden’s (1977) work. He noted that within the general class of fuzzyset tr:?ories (i.e., a set theory that conforms to (2.8) and (2.9)) there are alternatives to Zadeh’s characterization of fuzzy intersection and union. Thus, instead of the minim,Jm rule for fuzzy intersection, (2.3), we can use a multiplicative expression (due to Gougen, P969): ( v x E D) (CAnB (x) = CA (Xh (X)h And applying DeMorgan’s law to the above yields an alternative to the minimum rule for union, (2.4), namely,

D. N. Usherson and E. E. Smith

56

D) (cAu B(x) = CA(x) + cg (x) - cA (x)‘ca (x)). Oden (1977) showed that these alternative formulations provide a better account of subjects judgments of truthfulness of complex statements than do Zadeh Tformulations. But the above alternatives still yield almost all the contradictions we derived; e.g., as long as the two constituents of a logically empty set have characteristic function values between 0 and 1, their product must be greater than 0. Similar remarks apply to all the contradictions we derived, with the sole exception of that concerning the disjunctive concept of wealth (Section 2.4.3). (Referring back tc Table 1, given a reasonable choice of characteristic function values corresponc!ing to the various liquidity and investment amounts, the above expression for union correctly selects B as the wealthiest individual.) In short, we know of no alternative set theory that can be joined with prototype theory to account for all the evidence about conceptual combinations and truth conditions. Nor does it seem promising to fuzz the notion of truth to allow for partial falsification of thoughts, as was seen in Section 3.5. Perhaps a novel set theory and logic can be developed whose association with prototype theory will be free of the difficulties raised in this paper; such a development would be of considerable theoretical interest, in our opinion. Until a viable alternative to fuzzy-set theory materializes, however, prototype theory cariI;ot be suitably extended; and the possibility that no such alternative can be developed ought not to be minimized. A second possibility is to forget about prototype theory entirely. We count this as too extreme. For one thing, the empirical research that has resulted from prototype theorizing points to the use of similarity in establishing membership for many simple concepts. We mentioned some of the relevant findings in Section 1.l, and there are many other interlocking results of this kind: (see, e.g., Rosch, 1978). So prototype theory seems to capture something about the natural use of simple concepts. Furthermore, the leading intuition behind prototype theory - that concepts are often vague and apply to different objects to different extents - is quite cornpelling, and central to theories of reasoning. It is well known, for example, that the principle of mathematical induction cannot be reliably applied to vague predicates. To see this, define F to be the numerical predicate ( ‘tr

X E

grains of sand brought together do not constitute a heap. Note that F(O), and if F(k) then F(k + 1) (0 grains of sand brought together do not constitute a heap, and if k grains won’t do the job neither will k + 1). So mathematical induction leads to the false conclusion that no matter how large k gets, k grains of sand brought together do not constitute a h2ap. An

On the

adequacyof prototypetheor)-

57

adequate theory of concepts should provide an illuminating characterization of the class of vague predicates, and this is the goal of prototype theory. The third possibility is thrt prototype theory is by its nature incomplete because it is about only a lin.,;ed aspect of concepts. To make this clearer, we can distinguish between a concept’s core and its identffication prmedure; the core is concerned with those aspects of a concept that explicate its relation to other concepts, dnd to thoughts, while the identification procedunz specifies the kind of inform&ion used to make rapid decisions about membership. (This distinction is similar to one proposed by Milrer and Johnson-Laird, 1976).We can illustrate with the concept woman. Its core mi&t contain information about thz presence of a reproductive s/stem, while its identification procedures might contain information about body shape, hair length, and voice pitch. Given this distinction it is possible that some traditional theory of concepts correctly characterizes the core, whereas prototype theory characterizes an important identification procedure. This would explain why prototype theory does well in explicating the real-time process of determining category membership (a job for identifjcation procedures), but fares badly in explicating conceptual combination and the truth conditions of thoughts (a job for concept cores). This seems to us to offer the most satisfactory reconciliation of the present failures of prototype theory with its previous successes. References Bolinger, D. L. (1975) Agwcfs of Lmguuge, (2nd Edttion). New York, Harcourt Brace Jovanovieh. Cougen, J. A. l1%9) The logic of inexact concepts, Synthese, 19, 325-373. Hersch, H. M. and Caramazza,A., (1976) A !kzy set approach to modifiers ard vaguenessin natural language. J. #per. Rvchd.: Gen., IOS, 254-276. MUler. C. A. and Johnson-l&d, P. N. (1976) knguage and Perception. Cambridge,Hward Univcrsity Prew. Qdeu, G. C. (1977) Integratiai of firzzy logical information. J. exptx Pgvchol.: Hum. percep. pefwm,3, 565-575. PsBner,M. 1. and Keele, S. W. (1968) Qn the genesis of abstractideas. J. exper. Psych& 77,353-363. Potter, M. and Paulconer, B. (1979) Understandingnoun phrasca,J. verb. I&e@%vub. Behav., I& (5). m9=522, Reed, S. K. (1972) Pattern recognltlon and categorization, Cw P@‘chsl., .%382407. Rmb, N. (1%9)Many=wiuedlo~cs, McGrawHiU. Rips, L, J., Shoben, E. ,I. and Smith, Z;.E. (1973) Semantic distance and the verification of semantic relations. J. verb Lcucm.verb. khav,, lc?, l-20. Rosch, E. (1973) On the inter&4 structure of perceptual andeemanticcategories. In T. E. Moore (Ed.) Co@Mve development and the acquisitim of lhguag~. New York, Academic Press Rosch, E. (1978) Principles of categorization. In E. Roscl: and B. B. Lloyd (Eds.) Cognition und cutegotiutlon. Potomac, MD, Erlbaum.

58

D. N. Oshersonand E. E. Smith

Rosch, E. and Mervis,C. B. (1975) Family resemblance: Studies in the internal stntctum of categories. Cog. Pwchol., 7,573-605. Rosch, E. H., Simon, C. and Miller, R. S. (1976) Structural bases of ty&ality effects. J. expefi Pqy= chd.: Hum. Percep. Perform., 2,491-502. Smith, E. E., Shohen, E. J. and Rips, L. J. (2974) Structure and process in semantic memory: pi featural I,todel for semantic decisions. Psychd. Rev., 81,214-241. Tvenky, A., (1977) Features of similarity, Psychol. Rev., 84,327-352. Zadeh, L. (1965) Fuzzy sets.hform. confrcd, 8, 33S-353. Zadeh, L. A. (1975) Calculus of fuzzy restrictions, in L. A. Zadeh, K. Fu, JC.Tanada and M. Shimura (Eds.), I;trzzy se& and their applications to cognitive and decision processes, Academic Press.

La thCorie du prototype considbre qu’il existe des de& d’appattenance i l’extension d’un concept dktermink par ?a similitude avec le “meilleur” exemplaire de ce concept IOUpar quelqu’autre mesure de tendance centrale). Cet article envisage la compatibilitd de cette proposition avec deux critkes d’adiquation concernant * s thiories des concepts. Le premier critke conceme la relation entm les concepts complexes et leers contribuants conceptuels. La seconde a trait aux conditions devirit6 pour les propositions portant sur les inclusions simples.

CognJriof#,9 (1981) 58-72 @ ElsevierSequoia S.A., Lausanne - Printedin the Netherlands

GILLtAN COHEN” Univtmity of Oxfonl

Abstract Two experiments examined the effects of aging on the kind of inferential reasoningrequired tn ~~rnpr~h~nding discourse. In Experiment I old subjects made more errors than young subjects in soivinglogicalproblems framed in everyday language. Unlike the young subjects they had more difficulty when the problems were spoken than when they were written. Experiment 2 revealed that old subjects are inefficient at extracting implicit information during reading; they fail to generate bridg’ng inferences to supply missing information, so that comprehension is restn*cted to explicitly stated information. The results show that verbal reasoning ability is impairedin old age and that this affects language comprehension in both listening and reading although the deficit is more marked in listening, Introduction

Inferential reasoning is an integral part of language comprehension, and recent evidence has shown that it is vulnerable to the effects of aging. Although it is commonly asserted that ability is well preserved in old age, Cohen (1979) and Zelinski, d Schaie (1979) reported marked ag+related deficits in the comprehension c+fspoken discourse. Cohen’s study when old people listened to short spoken passages relat in ability to answer questions everyday events they were clearly de& requiring inferences fromtic? facts they heard, In contrast, their abilitv to answer which required simple reproduction of explicitly stated facts was d. The old group of aubjjactadid as well m a matched group on Verbatim Questions, but had much more difficulty with ce Que&ions. The furding that old people typically fail to draw inferences correctly is of conddenble importance since in everyday interchanges, both spoken and written, much informaa.on is imptied rather tw being explicitly stated. *IW rwwch was ~luppart@d by a grant from the sodal Scfanos Rwarch Cbmcil. Reprint requests should be sent to Gillian Coha~~,Dcpartmcnt of E%psrbncntalFaych&#y, Univerdty of Oxford, South ParkaRoad, Oxford OX. 3WD,UK,

A0

GiiIianCohen

Numerous studies (e.g., Bransford, Barclay and Franks, 1972 ; and Frederiksen, 1975) have demonstrated that inferences are normally constructed in the course of comprehension, and the memory representation of a text includes both explicit and implicit information. There is more tbaunderstanding than meets the ear, and failure to process implications will result in limited and impoverished comprehension. Consequently, it is of both practical and theoretical importance to determine exactly how, why, and in what circumstances comprehension is defective in old age. There are several different factors which may be contributing to the old people’s difficulty in answering inference questions. ( 1) Impaired recall

Qld people may construct inferences at the time of input, but forget the inferences they have constructed. Alternatively, they may neglect to construct the inferences while they are listening to the input and forget some of the facts they have heard, so that they arc unable to construct the inferences later in respoase to the questions. The first of these alternative explanations is unlikcy to be correct. Since the explicit facts required to answer Verbatim Questions are well retained there is no reason why inferences, if they have once been constructed, should be forgotten, especially as th6 deeper level of proa:%mg required to generate inferences generally ensures better retention (Craik and Lockhart, 1972). The second alternative is not ruled out. Ketrospectivc construction of inferences when the question is posed cQrzld bt handica,pped by impaired recall because, while a Verbatim Question requires recall of a single explicitly stated fact, delayed inference-making requires recall of more than one of the previously presented facts. The ability to answer Inference Questions correctly will therefore be adversely affected by loss of information to a greater extent than the ability to answer Verbatim Questions. If this is so, the age-related deficit in inference-making should disappear when the task is not memory dependent, and all the information is available for review. (2) Reduced lxxessing capacity A pervasive feature of mental performance in old age is a slowing down of

the rate of information processing. Understanding speech demands continuous perception, analysis and storage of a rapid input while at the same time successive propositions must be integrated and inferences constructed. There may be a tendency to omit inference construction when the rate of input exceeds the rate at which all these processes can be carried out. This sugges-

hferen t&Ireasoningin old age

61

tion receivo,d some support in Cohen’s 1979 study since the old subjects made more errors on Inference Questions with a faster rate of speech input, while Verbatim Questions were unaffected by increased rate. If rate of input is crucial, then the age related deficit in inference-making should be less evident in self-paced comprehension of written material. (3) Impaired reasoning ability A further possibility is that the logical reasoning ability required in inference making becomes deficient in old age. Tests of nonverbal reasoning ability such as Raven’s Progressive Matrices are known to show pronounced effects of aging and there is some evidence that verbal reasoning may also be affected (Horn and Cattell, 1966). Reasoning depends on the ability to manipulate, recode, or reorganize information in working memory. Numerous studies (e.g., Talland, 1968) have reported a related defic’l in shori term memory tasks such as backward and running span which require reorganization of the input. Inference construction which relies on integrating and cross-relating different propositions in working memory might therefore be expected to show an analogous deficit, and age differences should be more marked in inferential reasoning tasks which involve more complex transformaticns. It is quite possible that all the factors suggested above may be contributing to thJzrobserved deficit in infel-ence making. The experiments which follow were designed to try to determine the relative influence of each of these factors in inferential reasoning in old age.

Experiment 1 Logical reasoning with spoken and written problems. The aim of this experiment was to compare old and young subjects’ performance on logical reasoning problems with written and spoken presentation, since it is of practical importance to establish whether information is easier for old people to handle when it is written, or when it is spoken. For the written presentatiou unlimited study time was allowed so 9s to eliminate any effect of reduced rate of processing. Half of the written problems were presented in such a way as to involve no memory load. Comparison of performance under these different conditions of presentation allows the importance of rate of input and retention to be assessed, and any residual deficit in old people’s performance when rate of processing is unconstrained and no memory load is imposed can be attributed to diminished reasoning ability.

GillianCohen

62

Method Subjects

Two groups of subjects were tested. A group of 30 old subjects had a mean age of 69 (range 65-79). They were retired professional people who had all received university education or professional training. They were in good health and came into the laboratory for testing. Two pre-t%ts were administered to each subject, the auditory digit span test and the vocabulary test from’the Wechsler Adult Intelligence Scale. For the old group the mean digit span (forward + backward) was 13, and the vocabulary score was 69 points out of a maximum of 80. . A group of 30 young subjects, mean age 23, age range 19-29, were matched as closely 3s possible to the old group for educational level. The young group had a l-neandigit span of 14, and a vocabulary score of 69. The pre-tests served as a basis for matching the groups, and to establish that hearing was adequate for the experimental tasks, and memory ability within the normal range. Materials

Fo?y problems were constructed. Each consisted of a few sentences which constituted the premises, the mean length of each problem being 35 words. Half of the problems were followed by a conclusion to be evaluated (C problems), and half were followed by a question to be answered (Q problems). The premises were Framed in simple everyday language and represented five different types of inference. Examples were: 1.

Modm Ponens.

“If the man was lying when he said he was in Birmingham all day on Saturday then he could have been the murderer. The police checked his story and foijnd he had never been in Birminghamat all”. Conclusion: Ure man could have been the murderer. 2. Modus Tollernlrso Tollens. “If Bill played bladly in the football match he would lose his place on the school team. After the match the captain told him he would stay on the team”. Conclusion: Bill had not played badly in the match.

Inferentialreasoningin old age

63

3. Affirming the consequent.

“Going on holiday abroad with the whole family is very expensive. Harry took his wife and children to Spain for two weeks and it cost him a lot of money. John’s family holiday was also expensive”. Question: Did John take his family abroad for their holiday? 4. Denying the Antecedent. “If someone is a good bridge player he is bound to find whist boring. We tried to teach George to play bridge but he was hopeless”. Conclusion: George does not find whist boring. 5. Linear Syllogism. “Jean runs slower than Sue, and Pat does not run faster than Jean”. Question: Does Pat run faster than Sue? Twenty of these problems (4 of each type) were allocated to the Written condition and were typed in a booklet. Twenty matching problems were allocated to a Spoken condition and were recorded at a normal speech rate and played on a Sony TC 9% cassette recorder. In the Written condition each problem appeared on a separate page of the booklet. For 10 of the problems the question or conclusion was typed on the same page immediately below the premises (the No Memory condition). For the other 10 problems the question or conclusion appeared on the page following the prpmises which had therefore to be held in memory (the Memory condition). In the Spoken condition the speaker asked the question or stated the conclusion immediately after each set of premises. In both Spoken and Written conditions the: order of the different forms of inference and of C problems and Q problems; was random, but was the same for both conditions. Procedure Subjects were tested individually. Half the subjects in each group worked through the Written condition first, and the Spoken condition next, while half worked in the opposite order. In the Written condition the subject was asked to evaluate each conclusion and respond by circling one of three answers, True, False or Perhaps, and similarly to answer each question by circling Yes, No or Perhaps. Subjects were encouraged to take as much time as they liked, but were instructed not to turn back to check earlier responses. In the Memory trials, where the question or conclusion appeared on the page following the premises they were instructed not to turn back to study the premises ag&. In the Spoken condition subjects responded orally. After

64

GillianCohen

heming each question they were asked to respond, Yes, No or Perhaps, and

after each conclusitan True=,False or Perhaps. After each response the tape was restarted and the *rext problem presented. All the subjects took substantially longer to answer the written problems than the Spoken problems. The mean time taken to complete the Written condition was twice ae long as the time taken to complete: the Spoken condition. Scoring The response category of Perhaps was included because in everyday discourse the r&s of formal log& do not apply strictly. Conclusions which are logically invalid may be empirically possible. Hence in the examlple (4) above, although it is logically false to conclude that George does not fmd whist boring, it is empirically possible. Consequently in this example either False or Perhaps were scored as correct responses and only True constituted an error. By this criterion Perhaps is a legitimate response to all types of problem except the Linear Syllogisms.

ResUltS Table 1 shows the percentage of errors made by each group. Analysis of variance (Subjects X Groups (old/young) X Conditions (written, spoken) yielded main effects of Age (F = 19.8, df 1.58, p < 0.001); of Condition (F = 30.0, df 1,58, p < 0.001); and an interaction of Age X Conditions (F = 4.15, df 158, p < 0.05). Post hoc comparisons by Newman Keuls testc showed that the old group made signitic6antly more errors in the Spoken condition than in the Written condition (p -: 0.01). For the young group there was no significant difference between Spoken and Written conditions. The young group were superior to the old group in both Written and Spoken conditions (p < 0.01). Neither old nor young subjects showed any difference between Me,nory and No IMemory trials in the Written condition. The total ml ;lbers of errors Table 1.

Error percentagesfor old and young subjectson Spokn and Written Problems,Experiment I

Old Young

Written

Spoken

29% 20%

37% 22%

Inferentialreasoningin old age

65

for the Old group were Memory = 87, No Memory = 90, and for the Young group, Memory = 63, No Memory = 58. The two groups also did not differ in their tendency to reason empirically rather than logically as indicated by the fre luency of ‘Perhaps’ responses. The Old group responded ‘Perhaps’ to 23% of the problems and the Young group responded ‘Perhaps’ to 25% of the problems. The order of difficulty for the different forms of logical inference was also the same for both groups. Difficulty increased in the order Modus Ponens, Modus Tollendo Tollens, Linear Syllogisms, Denying the Antecedent, Affirming the Consequent. Both groups made more errors with C problems than with Q problems, because both groups found particular difficulty in rejecting false conchlsions. Table 2 shows the percentage of errors made on Q problems and in judging true and false conclusions in C problems. It is apparent that it is especially hard to disagree with Spoken false conclusions, although by x2 the distribution of errors did not differ significantly between Old and Young groups, (x2 = 0.8). Table 2.

Error percentages for old and young subjects on Q problem: and C problems in Experiment 1 _ Spoken

Written 0 Problems

Old Young

21.6% 13.6%

C Problems True

False

21.6% 12.5%

45% 36.6%

Q Problems

31% 15.5%

C Problems True

F&Z

19.1% 13.3%

59.4% 41.1%

Discussion

The findings that emerge from this experiment go some way toward clarifying the nature of the age related deficit in comprehension, The Old group, unlike the Young group, found it harder to make the correct inference when the input was spoken than when it was written. This can most plausibly be attributed to the differences in the rate of input. In the Written conditilon input was self-paced, unlimited study time being allowed; in the Spoken condition the rate of normal speech may have exceeded the processing capacity of old people so that they are unable to carry out proceF;es of inference construction concurrently with speech perception. The exptiriment gave no evidence that the deficit originates in forgetting since in the Written condition the Old

66

Gill&nCohen

group (jjd not make any .more elrOrs on trials when the premises had to be

held in memory. The retention interval is, of course, very short, spanning only the time necessary to turn the page and read the conclusion or question. This may be too brief a delay for memory shjortcomings to emerge. Nevertheless, it does suggest that the inferior performance in the Spoken condition is unlikely to be due %imply to the memory load imposed in listening to spzech. The retention interval while the spoken question or conclusion is presented is not longer than the retention interval for the memory trials of the Written condition. It could, of course, be argued rhat the results are reflecting a modality effect rather than a rate of processing effect. If auditory memory were inferior to visual memory in old age, then the observed deficit in making inferences from spoken inputs could be attributed to the modality of the input. However, the experimental evidence runs counter to this hypothesis. Memory iTor auditory inputs is usually found to be le% affected by aging than memory for visual inputs (Craik, 1977). Interestingly, the finding that Memory trials were no more difficul” than No Memory trials accords with recent results reported by Hayes-Roth and Walker (1979). They concluded that in memorizing texts, related facts are organized into conflgurations from which valid inferences can be read off, and that this process may actually be easier than searching back through a text which remains available for the relevant facts to justify an inference. On this interpretation, it seems more likely that in listening to speech old people do not have enou;Thtime to organize facts into configurations which have the logical relationship built into them. The fact that the Old group made more errors than the Young group even on the easier Written condition with processing time unlimited indicates that inferential reasoning per se is affected by aging. Logical reasoning with verbal problems appears to decline with age in the same way that logical reasoning with nonverbal problems such as Raven’s Matrices does. Although it was anticipated thirt old and young might show differential effects of task difficulty, this did not emerge clearly from the results. The order of difficulty for the different tyw= of logical inference and the magnitude of the differences between them were simiPar for both groups. Both groups found it harder to evaluate conclusions than to answer questions because both tended to accept false conclusions as true. In the two invalid forms of inference, affirming the consequent and denying the antecedent, illicit conversion of the premises caused errors. The relative complexity of different problem types may have been obscured by the tendency revealed in a substantial proportion of the responses to reason empirically rather than logically, and by the fact that these problems vary in the extent to which they seem ‘natural’ when expressed in everyday language.

Inferentialreasoningin old age

67

Experiment 2 Generating factual inferences. In place of the rather artificial logical problems posed in the first experiment, this experiment tested ability to construct inferences based on factual knowledge and so to generate the bridging assumptions that are needed to make sense of a text and to preserve its continuity and coherence. An example of this kind of inference is taken from Keenan and Kmts:h’s (1974) study of the comprehension of Explicit and Implicit texts. Explicit text

A carelessly discarded burning cigarette started a fire. The fire destroyed many acres of virgin forest. Implicit text

A burning cigarette was carelessly discarded. The tire destroyed many acres of virgin forest. In the implicit version the propositio;l that the cigarette started the fire must be inferred by the reader. Keenan and Kintsch conch&d from iheir results that implicit propositions are normally inferred by the reader while the text is being read. The present experiment compares the ability of old and young readers to fill out their comprehension of a text by spontanc: ~sly constructing these implicit propositions to bridge the gaps in the text. Reading speed for Implicit and Explicit texts was measured and compared, and the ability to answer Implicit Questions (relating to the implied propositions in the Implicit texts) was compared with the ability to answer Explicit Questions (relating to propositions explicitly stabId in the Explicit Texts). If old people are inefficient at generating this sort of factual inference as well as at logical reasoning they should have relatively more difficulty in answering Implicit Questions. A difference in reading speed for the two kinds of text might also be expected. If inferences are constructed, and this takes additional time, then the Implicit texts might take longer to read. If inference construction is neglected then reading speed for the two kinds of text should not differ. Method Subjects

Sixty subjects were tested, 30 old subjects and 3c) young subjects. They were the same as in Experiment 1.

Six stories were constructed. ‘T&e were Implicit texts and 3 were Explicit texts. Each story was typed on a separate card. The mean length of the Implicit stories was 55 words and of the Explicit stories 56 words. The stories recounted sinple events, aId were dissimilar in coc:ent (e.g., The Robbery, A Camping Holiday, The Accident) so that they were not likely to be confused. Procedure

Subjects *vereinstructed to read silently through each text as fast as possible, but were told to understand it thoroughly, bearing in mind that they would be asked questions about it later. The order of texts alto,rnated between Explicit and Implicit with half the subjects starting with an Explicit text and half w&h an Implicit text. The time taken to read each text was measured using an Advance Instruments TC 12 timer. The experimenter started the timer as each card was exposed, and the subject stopped it by pressing a key as he ftilished reading each text. Immediately after ersch story had been read the next was presented. After all 6 stories had been read, the questions were asked and answered orally. There were 4 questions relating to each text (24 questions in all), Texts were queried in the same order as they had been read, the questioner tiit identifying each story by its title (e.g., “The first story you read was about The Robbery...” ) and then asking the relevant questions. Questions urere openended Wh questions. The order of questions reflected the order of explicit or implicit propositions within the text. On average the delay between reading the relevant part of the text and answering the question was about 3% minutes.

Results and d%cussion Reading speeds and error percentages for the Old and Young groups are shown in Table 3. The error scores were submitted to analysis of variance (Subjects X Groups (old/young) X Conditions (Explicit/Implicit). Main effects of-Age (F = 13.5, df 1,58, p < O.OOl), and of Condition (F = 23.3, df 1,58, p < 0.001) were found. There was also a highly significant interaction of Age X Condition (F = 30.0, df 1,58, p < 0.001). Subsequent Newman Keuls tests showed that the Old group made significantly more errors on Implicit texts than on Explicit texts (p < 0.01). For the Young group there was no diffference. The Young group were superior to the Did

Inferentialreasoningin old age

Table 3

09

Error percentagesforold and young subjectsin Experiment 2. Weaditq.: speeds are shown in brackets

Old

Young

Explicit

Implicir

19.4% (290 w.p.m.1 13.6% (310 w.p.m.1

41.6% (263 w.p.m.1 12.5% (280 w.p..n.)

group on the Implicit text questions (p < O.Ol), but there was no difference between the age groups for the Explicit text questions. This parallels the finding from the study by Cohen (1979) where old people did a~ well as young people in answering verbatim questions, but were less good at answering inference questions. Analysis of variance was carried OULon the reading speeds. There was no effect of age group (F = 0.5, df 1.58, n.s.). Explicit texts were read faster than Implicit texts (F = 30.0, df 1.58, p < 0.001). There was no interaction of Age X Text type (F = I;.J, df 1.58, n.s.). Inspection of the data showed that querying information stated or implied early in the text as opposed to information later in the text followed different trends depending on whether the information wrasexplicit oi implicit. Figure 1 represents this finding. Analysis of variance (Age Groups X Explicit/Implicit X Position X Subjects) showed a significant interaction of Position X Explicit!Imphcit (F (lS8) = 25.9, p < 0.001. Although age group interacted significantly with Explicit/ Figure 1. Percentage emors in the recallof explicit and implicitfacts from the first half and the second half of the passages. n--aExpM CM

Facts

ImplIcIt Facts

100

J

50 10 Lo 30

w 2tldHdf 1stHolf YOUNG

P g? 2o 10 01

MHcltf

OLD

2ndM

70

GillianCohen

Implicit (F (1.58) = 27.1, p < 0.001) the triple interaction of Age Group X Explicit/Implicit x Position was not significant (I; (1.58) = 2.8, n.s.). FOR Explicit facts errors show a primacy effect increasing from the first half of the text to the second half. For Implicit information the opposite trend appears. Implicit queG:tionsare answered better from the second half of the text. A tentative interpretation of this finding is that memory for the information given in the tIext (the Explicit facts) suffers from increasing overload later inthe text so that performance declines; Implicit facts on the other hand, are extracted more efsciently later in the story when more information is available and a global memory representation of the whole story can be formed. Possibly subjects do not begin to construct bridging inferences early in the text because they expect the information to be forthcoming. When they realize that it is rniss~g they begin to make the necessary inferences. The difierences in direction of the slopes for Explicit and Implicit facts indicates that the inferences are not being constnrcted retrospectively from remembered information when the questions are posed. The fact that the slope for lmpEdt Facts is the same for young and old suggests that old people also attempt to construct inferences during reading but are much less efficient at doing so. The Old group were quite strikingly poor at answering questions related to implied information, while the Young group were as good, or even slightly better at answering Implicit questions than at answering Explicit questions. Both groups took more time to read and comprehend the Implicit texts, but, whiIe the Young group achieved a speed-accuracy trade-off which allowed them to maintain the same error rate as in the Explicit texts, the Olldgroup failed to reduce their reading speed sufficiently and their error rate increased dramatically in consequence. In Keenan and Kintsch’s study they did not fmd that Explicit texts were read faster, but their Explicit texts were cobsiderably longer than the Imphcit ones, and they meast-red total reading time rather than reading speed”in words per minute so that this difference in speed of processing could not be expected to emerge.

The conclusion that emerges from these experiments is that in old age there is a general de&it in inference-making affecting a variety of different kinds of inferential reasoning. There is an age-related deserioration in making inferences based on logical relationships (as in Experiment 1) and in making inferences based on factual knowledge (as in Experiment 2). ‘Ihe difficulty that 01d people have in drawing correct inferences is enhanced when the input is

Inferentialreasoningin old age

71

spoken, probably because the rate of input exceeds their processing capacity. Nevertheless, a deficit was still evident when the input was written and reading time was unconstrained and self-paced as in Experiment 2 and in the Hritten conditions of Experiment 1, indicating that verbal Ieasonmg ability is adversely affected by aging There is no evidence from these experiments to support the contention that the difficulty old pet; ’ 3 hav,: in making inferences stems from impaired retention of information in memory. In Experiment 1 the performance of old subjects did not differ for the Memory and No Memory conditions, although it could be argued that the retention interval in Memory condition was too short for memory impairment to exert an effect. However, in Elxperiment 2 the retention interval was much longer and the memory load greater. The old subjects still retained the Explicit facts almost as well as the i’oung subjects. If their failure to answer Implicit questions correctly was due to inability to retain the facts of the story well enough to construct inferences retrospectively when the questions were asked, then the slope for Implicit facts as a function of position in the text would have been the same 3s the slope for Explicit facts. The difference in direction of the slopes shows clearly that the ability to answer Implicit questions is not dependent on retention of Explicit facts. To summarize, old people have difticulty in extracting information which is implied but not stated. The inference construction stages of comprehension are either omitted, or carried out slowly and inaccurately. The difficulty is increased when the rate of input exceeds processing capacity. As a result, even for the well educated ar,d well preserved old people tested in this study, language comprehension tends to be restricted to atomic facts and fails to include information derivable from the relationships between them. References Bransford, J. A, Barclay, J. A. and Franks, J. J. (6972) Sentence memory: a constructive versus interpretive approach. Cog. Psychol., 3, 193-209. Cohen, G. (1979) Language comprehension in old age. Cog. Psychol., 11,412-429. Craik, F. I. M. (1977) Age differences in human memory. In J. E. Birren and K. Warner Schaie (Eds.), Handbook of the Psychology of Aging. Mew York, Van Nostrand Retiold Co. Craik, F. I. M. and Lockhart, R. S. (1972) Levels of processing: A framework for memory research. J. Y&, Learn. Verb. Behav., II, 617-684.

Frederiksen,C. H. (1975) Effects of context-induced processing operations on semantic information acquiredfrom discourse. Cog. Psycho& 7, 139-l% Hayes-Roth, B. and Walker,C. (1979) ConfiguraJeffects in human memory: the superiority of memory over external information sources as a basis for inference verification. Cog. Sci., 3, 119-140. Horn, J. I,. and Cattell, R. B. (1966) Primarymental ability factors.J. Gerontol., 22, 210-220.

72

GillianCohen

Keenan, J. M. amI Kintsch, W. (1974) Memory for information inferred during reading. In W. Kin&h, The Representation ofMeaning in Memory. Lawrence Erlbaum Associates, Hillsdale, N.J. Talland, G. A. (1968) Age and the span of immediate recall. In G. A. Talland (Ed.), Humpn aging and Behaviour. London/New Yorlr, Academic Press. Zelinski, E. M., Light, L. L. and Schaie, W. (1979) Age differences in memory for facts versus inferences. Paper presented at the Annual Meeting of the Psychonomic Society, Phoenix.

On a 6tudii avec deux experiences Ies effets de I’lge sur le genre de raisonnement infedrentiel requis

pour comprendre un discours. Dans l’experience 1 les suJets Bgtk font plus d’erreurs que las sujets jeunes en risolvant des probknes logiques present& en langage de tous les jours. Contrairement aux jeunes sujets ils ont plus de difficult& avec les probkes present& oralement que par ecrit. L’expkience 2 montre que les sujets fg& sont mefflcaces quand il s’agit d’extraire l’information implicite durant lalecture; ils kchouent B faire les infirences kcessaires pour supnIGer1 l’information manquante de telle sorte que la compr6hension est reduite a Ymformation donnke. Les Ssultats montrent done que la capaciti de raisonnement verbal est appauvrie chez les personnes Gg&eset que cela affecte la comprehension du langage 6 la fois dans l’ecoute et la lecture quoique ce deficit soit plus net dans I’&oute.

C0gnifion. 9 (1981) 73-114 @ Eseviet Sequoia S.A., Lausanne - Printed in the Netherlands

5

Building theoriesof reauingability: On the relation between individual differencesin cognitiveskills and readingcomprehension* THOMAS H. CARR Michigan State University

Abstract To date most theories of reading ability have emphasized a single factor as the major source of individual differences in performance. However there has been littleagreement on what thatfactor is. However, candidates have included visual discrimination, phonological and semantic recoding, short-term memorv, and utilization of linguistic knowledge and context. The singlefactor theories are summarized. Literature is then reviewed to show that no single-factor theory is likely to be right, because a very wide range of component skills and abilities has in fact been shown to covrelate w,‘th reading success, Among them are discrimination of letter location and letter order during perceptual recognition, use of orthographic regularity as an aid to visual code formation, use of spelling-to-sound regularity in phonological recoding, memory for word order, spontaneous identification cf syntactic relations, flexibility in prediction from syntactic and semantic context, and context-specificity in semantic encoding. It is concluded that more complex, multifactor models of reading ability are required, and some recent attempts to collect data conducive to such a model are described. In the process, three different approaches to id, tifying factors relevant to reading success are delineated. These are general abilities assessment, learning potential assessment, and component skills analysis, Two methods of conducting component skills analysis are presented, and it is recommended that they be used as converging operations, Finally, the results of a component skills analysis are used to construct a tentative example of a class of hierarchical modei’s of reading ability that can be pursued developmentally.

*I would hke to thank Roderick W. Barron,J. Kathryn Bock. Denise Frieder0x1, MaryAnn Evans, Margot Haynes, Janet Kistner, Martin H. Singer, Rose T. Zacks, and two anonymous reviewersfor their insightful comments and criticismsduring the preparationof this paper. Addresscorrespondence to Thomas H. Cbr, Department of Psychology, Michigan State University, East Lansing, Michigan, USA 48824.

74

Thotms H cm-r

Current theories and their viability

Psychologists always prefer a +nple theory to a complex one, as long as both explain the data. Consistent with this predilection, explanations of individual differences in reading comprehension and reading achievement have often focused on single factors. Vanation in one component skill has either explicitly or irrplicitly been heldhfisponsible for most of the variation observed m overall performance. Single-factor theories have been especially popular as explanations of severely deficient reading achievement or “dyslexia”. Such accounts would be grand in their parsimony were they consistent with all that is known about reading, but I will argue in this paper that single-factor theories are simply too simple to be correct. Current data indicate that no component skill can explain a sufficie;\t amount of individual variation in reading performance to warrent a single-factor theory. This suggests that a synthesis will have to be undertaken. There are basically four major clas%csof smgle-factor theories, hypothesizing deficits among poor readers in: (1) visual discrimination and visual code formation, (2) phonological and semantic recoding, (3) short-term memory, and (4) the utilization of linguistic knowledge and context to guide perceptual recognition and encoding. By defmition these hypotheses could not all be true if a single-factor theory were viable, and it does appear that they are not equally powetiul. How to rankorder them remains a matter for debate. Despite possible differences in explanatory power, however, the four classes of single-factor theories may each capture a piece of the truth, though indivjdual theories within some of the classes may be quite wrong. In order to support these claims, several bodies of literature will be reviewed vriirh the goal of demonstrating that a wide range of information processing skills has been shown to vary with reading ability, including most (though not all) of the skills around which single-factor theories have been built. The job facing researchers is therefore not to find out which single-factor theory is correct, but to discover how big a piece of the truth each of them actually captures. Some preliminary data pertaining to that question will be discussed. Because of the variety of skUls that have been shown to differ between good and poor readers, a plea will be made for theories that allow reading ability to be multiply rather than singly determined-research should be oriented towards the construction of a functional ath of reading-related individual differences in cognition, rather than toward identifjkg the-cause of ftiure at reading. The plan of the paper is to set the stage by summarizing the nature of the explanations for reading abihty offered in the major classes of single-factor theories. After a brief digression into the relative merits of terms like “dys-

Building theoriesof reading ability

75

lexia”‘, “reading disability”, and ‘*poor reading”, the range of information processing skills that vary with measured reading comprehension will be established. Where independence of a skill may be in doubt, an effort will be ICS??P, to show whether or not individual differences in that skill might better be regarded as the result of individual differences in some other more fundamental or *moregeneral skill. Frequent reference will be made to short-term memory explaklations in this regard, since short-term memory is the broadest and most flexible of the various concepts on which single-factor theories have rehed. Finally a few recent studies will be mentioned in which readers have been administered batteries of information processing tasks spanning large portions of the range of skills on which good and poor readers have been shown to differ. From these studies some conclusions will be drawn about the likelihood that we c(n.nsuccessfully establish the relative contributions of each skill to overall comprehension performance and identify patterns o. coherence and dissociation among the skills. If those things could be accomplished, we might be able to go on to identify subgroups of good or poor readers who show different profiles of skill strength and weakness, and who would therefore require different types or instructional arrangements in order to become better readez than they currently are. Single-factor theories of reading ability The first class of single-factor theories postulates a deficit among low-ability

readers in visual discrimination (Benton, 1962; Critchley, 1964, 1970; Orton, 1966). Traditionally these theories have focused on the perception of spatial orientation with the occurrence of reversal errors in letter identification, such as mistaking b for d or q for I”, serving as a major source of their support. Several studies now indicate that reversal errors might be more common among younger than older children, but are no more common among poor than good readers (Firth, 1972; Goodnow, 1972; Shankweiler and Lieberman, 1972; Vellutino, Steger, and Kendal, 1972). Visual discrimination theories therefore appe:-r to be unique in that the factor supposed by many of them to underlie the entirety of dyslexia may not account for any of it. However, the case is not closed on a relation between reading disability and some other characteristics of visual discrimination, including spatial location, spatial order, and temporal resolution (Mason, 1980; O’Neill and Stanley, 1976; Stanley and Hall, 1973). Visual disc,rimination theories of the kind just described have been concerned for the most part with primary processes of sensory registration, feature extraction, and feature combination. Several recent investigations have reached the conclusion that these primary processes are relatively un-

related to reading ability, but that poor re;: iers show less benefit than good readers from secondary feedback mechanisms in the visual system. These feedback mechanisms take advantage of predictable spelling patternsreferred to as orthographic regularity-to facilitate visual encoding (Frederiksen, Note 1; Mason, 1975; Massao and Taylor, Nots 2). Research on the utilizatron of orthographic regularity has not yet produced a singJe-factor model in which the major cause of reading disability is held to be a deficit in visual code formation resulting from inefficient feedback mechanisms. Given the history of reading research, that may be only a matter of time. However, problems in dealing with a closely related structural property of the writing s/stem have figured prominently in single-factor explanaZions of poor reading performance. This second class of single-factor theories postulates a deficit in phonological and semantic recoding. Acpordinlb to the recoding theories, word recognition is severely impaired because the poor reader has trouble translating from spelhngpatterns to pronunciation and to meaning. These theories vary tremendously in scope and degree of elaboration; four examples will illustrate their range. Vellutino (1977) argues that poor readers suffer a pervasive “verb91 deficit” that interferes with any and all activities requiring verbal labels to be activated or manipulated. Rozin and Gleitman (1977) on the other hand, severely restrict the domain of verbal abilities in which there are thought to be deficits among poor readers. As long as the relation between printed and spoken versions of words can be handled at the level of the whole word or the sy?!able, learning to recognize words is easy. But when word sounds must be analyzed into phonemes, word recognition becomes a difficult task to master. Because phonemes are abstract categories (Chomsky, 1970; Chomsky and HaJle, 1968; Rozin and Gleitman, 1977), spslech sounds and spellings do not correspond in one-toone fashion at the level of the phoneme. The ability to analyze speech sounds into phonemes and to make c~nntions between the abstract sound categories and spelling patterns is called “phonemic awareness” (Golinkoff, 1978)~and according to Rozin and Gleitman poor readers lack sufficient phonemic awareness to be able to learn to use phonics rules to translate print into spoken words and vice verse with any degree of facility. Em phonemic awareness hypothesis is a claim about the crsnsciously apdble linguistic knowledge of the poor reader (Rozin, 1976). A recoding theory of sm@ar scope comes from Perfe tti and Lesgold (1978), but focuses on procedural rather than declarative knowledge of spelling-to-sound correspondence In tirisview, the extent to which spelling-twound translation rules have been integrated into well established procedural subroutines or pronuntliation schemas is crucial to reading ability. Poor readers often have

Building theoriesof reading ability

77

much more difficulty than good readers in pronouncing unfamiliar but orthographically and phonetically regular letter combinations (dop, spod) yet manage to do quite well at pronouncing familiar combinations with which they have had considerable prior reading experience (dog, snot). Perfetti and Lesgold agree with Rozin and Gleitman that for “dog” a pronunciation is stored in memory and can be retrieved directly via paired-associate mechanisms, while for “dop” a pronunciation must be constructed via analytic spelling-to-sound translation rules. At some point the translation mechanism must deal with correspondences between phonemes and individual letters or letter clusters. Whereas Rozin and Gleitman are concerned with the difficulty of grasping the concepts underlying phonemic analysis, Perfetti and Lesgold are concerned with the difficulty of applying the concepts once they are grasped. Early in learning to read, carrying out spelling-to-sound translation demands a lot of attention according to Perfetti and Lesgold. Given that that capacity within the information processing system is limited, attention devoted to recoding is not available to support higher-order memory maintenance and comprehension processes. Recoding mechanisms become progressively more automated with practice. Until recoding reaches P sufficient level of automaticity, however, comprehension processes will suffer due to a lack of cognitive resources to be allocated to their operation. Poor reac rs have not been able to automate recoding mechanisms to the same extent as good readers, and as a re:sult do not comprehend text as well. This kind of argument was first advanced by Bryan and Harter (1899). Huey (1908) relied upon it as well, making Perfetti and Lesgold’s version one of the latest in a venerable line. Both the phonemic awareness and the attention demand hypotheses distinguish between recoding relatively unfamiliar words (which is thought to be hard) and recoding familiar words for which phonological and semantic codes are already stored as visually-addressable units (which is thought to be much easier). Jackson and McCleiland (1979), however, do not spare poor readers even the ability to utilize paired-associate mechanisms. In what might be called the higher-order access hypothesis, Jackson and McClelland argue that poor readers are slower and less accurate in activating any nonvisual internal representation of a printed word, including codes that are dready stored and need only to be retrieved rather than constructed. The higherorder access hypothesis is genericaIly similar to Vellutino’s notion of a verbal deficit, but is nrore firmly grounded in a theory of underlying mental operations. Because a particular operation, memory retrieval, is specified as the locus of the poor reader’s difficulti], Jackson and McClelland’s hypothesis is better def ;ed and therefore more testable than Vellutino’s (see Singer, 1979).

78

ThomasH. Cam

A third

class

of single-factor theories does not deny that poor rea&:rs have

trouble with phonemic recoding, and especially with rule-goverr:::d spelling-

to_scand translation, but SW recoding difficulty as a by-product of a more fundrmental deficit in short-term meml3ry. In many ways these theories have the most a priori power of the single-fat tor explanations, because “short-term memory” is a broad and malleable enough concept to potentially subsume a very large and diverse collection of idhiditd phenomena. For this reason proponents of short-term memory theories bear a particular burden of precision in defining constructs as well as in specifying exactly how the constructs are to explain the observed characteristics of reading performance. There are two main types of short-term memory theory that differ in the extent of the proposed storage deficit, and a third hybridized theory that demonstrates how short-term memory and recoding hypotheses can shade into one another. The broadest of the short-term memory theories might be called the generalized short-term maintenance hypothesis. Jorm (i979), for example, claims that all dyslexia, both developmental and traumatically induced, results from a relative inability to keep several codes or chunks of information simultaneously active in short-term store, and that a malfunctioning left inferior parietal lobe is responsible for this deficiency. A more limited short-term memory theory holds that the ability to maintain the identities of a series of items does not vary much with reading skill, but maintaining the order in which the items occurred gives poor readers much difficulty (Doehring, 1968; Torgeson, 1978). Corkin (1974) believes that this notion has great power to explain the inability of dyslexics to learn to read, and Singer (1979), though not proposing a single-factor model, argues that the serial order hypothesis has considerable merit. The third, hybridized, short-term memory theory is an extension by Perfetti and Lesgold (1978) of their attention-demand recoding hypothesis. Because of difficulty with recoding operations, poor readers are inordinately slow and inefficient at encoding new items into short-term sfort!, and they are also, according to Perfetti and Lesgold, slow and inefficient at clearing old items out of it. A smooth flow of information through shorlt-term store keeps comprehension mechanisms that make use of that information operating at their best. The hfficient encoding and forgetting of the poor reader causes a backlog or traffic jam in short-term store that interferes with comprehension. This backlog, in which accurate encodings cannot be made available to comprehension Processes at a fast enough rate, is called “hysteresis’” by Perfetti and Lesgold, from the Latin term meaning “to lag behind”. If hysteresis is severe enough, comp=bension processes can grind nearly to a halt and understanding can break down completely. Because of its hybrid character, the hysteresis hypothesis is hard to classify as either a short-term memory theory or a

Buildingtheoriesof reading ability

79

recoding theory. The locus of comprehension failure is considered to be short-term memory, but storage capacity per se is not the culprit. Trouble arises in getting codes for words into short-term store to begin with, and in forgetting them after their information has been integrated into a longer-term representation of the text of which they are a part. The hysteresis hypothesisis complicated, and Perfeili and Lesgold actually provide no evidence that codes are slow to be cleared out as well as slow to be entered into short-term store. One might even argue that their position contradicts the other short-term memory theories, in which forgetting is faster rather than slower among poor readers On the one hand, a kind of active or directed forgetting may be a part of comprehension-when constituent boundaries are reached, short-term store appears to be cleared of phonological codes representing surface structure in P way that ordinary processes of decay and interference (cannot explain (Chang, 1980; Clark and Clark, 1977; Davidson, 1978; Jarvella, 1970, 1971). If such a short-term memory mechanism exists, then individual difference; in its efficiency could affe-t reading comprehension. On the other hand, the hysteresis hypothesis could be reworked into a somewhat simpler claim without sacrificing the spirit of Perfetti and Lesgold’s argument for a close connection between recoding and memory maintenance. If the same source of limited capacity is needed both to recode visual material and to store the codes so activated, then the harder it is to recode the less can be stored (Baddeley, 1979; Baddeley and Hitch, 1974; Carr, 1979). Trading off storage for work, then, would reduce comprehension by increasmg short-term memory loss among readers who find recoding extraordinariIy difficult. This is in fact the attention-demand recoding hypothesis already attributed to Perfetti and Lesgold, with short-term memory specified as the particular higher-order process from which recoding steals resources. The three classes of models discussed so far all point toward bottom-up or data-driven processes as the source of reading disability--poor readers are forced at some point in the reading process to work with lower quality visual, phonoIogica1 or semantic encodings of the printed words than good readers. The fourth class of single-factor models places the blame for reading disability on a failure of top-down or knowledge-driven processes instead. Poor readers do not make as extensive or efficient use of world knowledge, of linguistic knowledge, or of the context provided by what has already been read to guide the acquisition and interpretation of data from upcoming parts of a passage. Two roles are usualIy attributed to top-down processmg, and either can be the main source of reading problems depending on the particular theory. First, increasing knowledge of the language shows print to be encoded in

80

nlomas fi. GUT

targer and larger units-- letter clusters rather than letters, whole words rather than letter clusters, constituent phrases rather than words (Laflerge, 1979; With, 1971, 1973). Greater unitization leads to gloater efficiency of encoding, short-term maintenance, comprehension, ;?d long-term storage by reducing the number of chunks of information required to represent a given text. Secon’d, greater sensitivity to the context provided by already-read text allows moire and better predictions about information that is likely to be coming up next. Greater context use leads to greater efficiency by turning reading into a !;ophistic:ated guessing ganie in which much less stimulus information has to be processed to gain a given level of understanding (Goodman, 1969, 1973; Rumelhart, 197’;, 1978; Friedrich, Note 3). Top-down deficit theorie;, then, postulate that poor readers fti to use linguistic knowledge to increase unitization or fail to use context to increase predictive support for perceptual recognitim and comprehension. 4s Rozin and Gleitman (1977) have summarized the arguments o$ the top-down deficit theories, poor reader. _4 ““plodders” mired in the data, processing each bit of print to the same depth .3ne bit at a time. Good readers are “explorers” who pick and choose, processing some parts of the text to great depth and other parts only superficially depending on the redundancy and predictability of each part’s content 3n relation to the overall meaning the text is conveying. On terminology

Now thst a summary of the single-f&or theories has been completed, the main thsk is to review the information processing literature in order to demonst-ats which of the skills addressed in all these theories do in fact vary with overall reading performance. First, however, a question of terminology ought to be raised. In summarizing the single-factor theories, I have used the terms “dysl~exia”, “‘reading disability”, and “poor reading” or “poor reader” ~4th apparent indiscrimination. This is true of the reading literature in general, and I would like briefly to discuss the, imphcations of adopting one terminology over another. Particularly significant consequences attach to the use of the term “dyslexia”, which will be criticized in the next sectilon. What is “dyslexia “? The tern “dyslexia” is usually used to refer to the reading performance of

people “who, despite conventional classroom experience, fail to attain the language &ills of reading, writing, and spelling commensurate with their intellectu:d abilities” (Jorm, 1979, p. 19). Though widely accepted since it was introduced by Critchley (1964), this dzfmition causes many more

Building rheoriesof reading ability

81

prcblems than it solves. First, reading as an input process and writing as an output process have somewhat different control structures, at least in proficient adults. Frith (1979) reports that the largest number of errors in word recognition bear visual similarity to II;&target word bur the preponderance of errors in spelling bear phonological similarity. If input and output processes operating on text are at least partially independent, then a person could be skilled at one but not the other. This actually happens according to Gibson and Levin (1975), who argue that good readers are not necessarily good spellers. Second, reading comprehension appears to be better established in nearly everyone than is writing composition (more people can read, understand, and enjoy William Faulkner than can write as good a short story). Thus reading, writing, and spelling don’t always vary together, even though recognizing and spelling words share some processes in common (Baron, 1979; Barron, 1979; Carr and Evans, in press, Simon and Simon, 1973) as do comprehending and composing prose (Bock and Irwin, 1980; Kintsch and van Dijk, 1078). If writing 2nd spelling were dropped and readir; .were used as the only criterion p?sformance, there would still be a good deal of doubt about whose reading performance is to be explained. This is because the remaining terms of the definition are themselves very difficult to define. First we have to decide what is meant by “conventional classroom experience”. For example, Evans (1979) has compared 10 classrooms in a large Canadian school system that use a formally organized, teacherdirected (or “traditional”) curriculum to 10 classrooms with similar students from the same school system that use an informally organized, student-centered (or “open”) curriculum. A larger proportion of students were reading below grade level in the infomral curriculum than in the formal curriculum. Should the informal classroom organization count as conventional experience, and should children from the two kinds of curricula be studied together? Or should the informal curriculum be excluded from the realm of conventional classroom experience? If it is excluded, then principled grounds for exclusion would have to be defined, and from then on only children who fail to learn to read in some restricted subset of the many school curricula in use around the world could be stludied as “dyslexic”. Other difficulties arise from the phrase “commensurate with their intellectual abilities”. The proper meaning of “comn,znsurate” and “incommensurate” can’t be determined scientifically. Guidelines can be agreed upon, but in the end the decision is largely arbitrary. Could we say, for example, that “incommensurate” reading ability is a grade-level competence two standard deviations below the grade-level equivalent of the child’s mental age on an accepted test of intellectual ability? If so, why? If not, why not?

82

77tamasH. Gm

And if so, what test of intellectual ability should be accepted? Ideas on the measurement of atelligence are in tremendous flux (e.g., Resnick, 1976) and it is unclear that the IQ test will remain a viable instrument even in cases where cultural bias is not a serious problem. An alternative to IQ testing that is gaining favorrapidly is the assesment of basic information processing skills (Carroll and Maxwell, 1979; Simon, 1979; Sternberg, 1977). Many of the basic skills to be assessed are implicated by single-factor theories as causes of dyslexia- if this approach were accepted, then equating readers in “intelligence” would create a paradox for the P’heories. But enough. Differences in reading achievement associated with differences in curriculum could be investigated directly by appropriate selection of subjects (e.g., Carr and Evans, in press) or else factored out statistically, and a large Jattery of converging indicators could be applied to give at least a reaonable feeling that children chosen for study as poor readers are not poor at everything. The sorts of primarily methodological refinements proposed so far, ho-wever,would not entirely meet my objections to the use of the term “‘dyslexia”. Although the defmition is relative, its applicaticzr?all too soon becomes absolute in the arguments of most people who employ it. Speaking of “the dyslexic” brings to mind a dichotomous distinction: you either have “dyslexia” or you don’t. A quick look at the subjects in studies cited by almost anyo.\e reviewing the reading literature indicates that this is not true, that there is n3 uniform set of criteria in use that will consistently identify one person as “dyslexic” and another as “normal”. As an illustration using some commonlycifted stud,” i-s, many of the subjects from the low ability group of Guthrie (1973) or Katz and Wicklund (1972) would not be classified as poor readers by Rozin, Poritsky, and Sotsky’s (197 1) standards, and few of Boder’s (197 1, 1973) clinical patients would have shown up in any of the available experimental studies- they were too bad. Yet some remarkably consistent differences tiave emerged from comparison of better and poorer readers regardless of where on the continuum of measured ability their absolute comprehension performances fall. Rather than discus&g “‘the dyslexic”, then, it might be more protitable at this stage of the game to Esk whether or not individual differences in specific information processing &ills are reliably related to individual differences in reading ability (see Calfee, 1977; Carroll and Maxwell, 1979; McClelland and Jackson, 1978). This question focuses attention on the fact that reading ability v&es continuously rather than discontinuously or bhnodally. Like distributions of tested IQ, distributions of tested reading ability are essentially normal, with a small bump at the low end. The bump in IQ r’,isttibutions represents people with congenital brain damage or chromosomal disorders, to which a medical model of mental retardation as a dichotomous condition can be applied.

Building theoriesof reading ability

83

The remainder of tht distributionisnot at all subject to dichotomous classification into a “retarded” group and a “normal” group. The same situation seems to hold for reading (e.g., Yule, Rutter, Berger, and Thompson, 1974). Therefore I will refer in the rest of the paper to better and poorer readers, or to good and poor readers as a shorthand, intending to indicate a contrast between two points located somewhere on the continuum of reading ability or achievement as assessed by comprehension tests. Depending on the study being cited, these points may be separated by anywhere from one to four grade levels or two to four stanines of the ability or achievement distribution observed at a particular age. This choice of terminology is not so much intended to add information that the use of “dyslexic” and “dyslexia” obscures, as it is to discard excess connotational baggage in hopes of finding a more neutral-and more accurate-way of referring to variation among readers in overall proficiency. Skill differences among readers

The next several pages will examine the evidence on information processing skills that vary with comprehension performance. I noted earlier that care has to be given to establishing the independence of these various skills from one another--to establishing that some particular kmd of cognitive performance is not just a special case of some other more fundamental or more general kind of performance. For that reason I will start with the evidence on short-term memory differences as a function of reading ability, since, as already noted, “short-term memo-y” is the most flexible of the theoretical constructs that have been used to explain limi ,;;ons on performance in reading-related information proceqsing tasks. Are short-term memwy differ.mes related to reading ability? The most encompassing of the short-term memory theories is the generalized

short-term maintenance hypothesis offered by Jorm (1979). The major emhat “dyslexic” children have a deficit in pirical claim in Jorm’s paper short-term memory for both visually and auditorially presented materialthat is, that reading ability and differences in short-term memory performance are reliably related. In the literature, several cognitive functions have been attributed to short-term memory, including temporary storage, executive control, and selective attention (Baddeley and Hitch, 1974; Carr, 1979). It is the storage function that Jorm considers important. This claim that reading ability and short-term memory performance correlate appears to be

true. However, when the claim is differentiated with respect tP what kind of information is being stored, items or their order of occurrence, it becomes apparent that the generalized short-term maintenance hypothesis is too broad to be useful. Substantial evidence now exists that item and order information must be considered separately in studies of reading-related information processing.’ In at least three different kinds of short-term memory tasks, order information is lost more rapidly than item information and the difference in rate of loss for the two kinds of information is greater for poor than for good readers. The tasks are reproduction (Corkin, 1974, Mason, Katz, and Wicklund, 1975), successive same-different matching (Singer, 1979; Singer, Lappin, and Allen, Note 4, Tallal, 1980), and delayed recognition with tachistoscopic stimulus presentation (Manis and Morrison, Note 5). In fact, memory for order is the major thing distinguishing poor from good readers in these tasks. The reproduction and same-different matching data show very little difference as a function of reading ability when item identities have to be remembered but their order of occurrence does not. In the delayed recognition task, poor readers are worse at both kinds of memory, but the disadvantage is larger when order is important. Thus a more accurate claim would be that reading ability is related to the ability to maintain ords over time, as proposed by the more specific serial order hypothesis. Since Sihger’s successive matching experiments showed similar results for visual and auiditory sequences, this relation May hold for both spatial and temporal order;. How might the ability to deal with these two kinds of order figure in the reading process? ci Sputial order and reading

Maintaining spatial order could potentially be important to reading at two levels. Readers might have to keep straight the spatial arrangement of letters during word recognition and they might have to keep straight the spatial arrangement of words or phrases during sentence comprehension. Data on eye movements by skilled adult readers (Rayner, 1978) indicate that useful visual information is usually obtained from a rather small area during each fixation, some 3 to 4 character spaces to the left and 8 to 10 spaces to the ‘Crowder (1979) has recently argued that in cases where the stimuUare already known in a generic sense, such as familiarwords OXletters, item and order information are the same thing but on different scales or at different levels of hierarchy. For example, item information requiresone to know that X occurred in List 3 but not in Lists 2 or 4, while order information requiresone to know that X occurred in Position 6 of List 3 but not in Positions S or 7. Even if item information does turn out to be reducibleto orderinformation for some theoreticalpurposes, however, the distinction is still important to readingif people show differential facili@ with information at the two levels of the hierarchy.

Building theoriesof reading ability

85

right of the fixation point. The average saccade length is about the same distance to the: right, 8 to 10 spaces. Tayior (1965) found that the mean number of fixations per 100 words of text doesn’t go below 100 until tenth or eleventh grade (average age 16 to 17 years) and is still around 90 among college readers. Data like these imply that people often fixate most of the words they read, especially after regressions are taken into account. As a rough approximation, then, spatial order is probably most important in the recognition of individual words. Multiword constituents involve the concatenation of information gained from successive fixations, confounding the need to maintain spatial order with the need to maintain temporal order. The major question regarding a relation between spatial order and reading ability is therefore concerned with the role of spatial order in word recognition.* Estes (1975) showed that in the immediate report of briefly presented letter strings, errors of location or transposition are common, such as saying that P occurred in the second rather than the first position of the string PJMRL. The memory requirements of tachistoscopic report are not especially large, since the strings are usually of word length and report usually begins within a second after the stimulus has been removed from view. Moreover, transposition errors occur even when the subject is cued to give only the letter that appeared in a particular position. On this basis one might begin to bonder whether discrimiibating order initially rather than maintaining it afterwards could be a difficulty in processing letter strings. Chambers (1979) discovered that people performing the lexical decision task took longer to reject nonwords that differed from a word by the order of two adjacent letters than nonwords that c :uld not be turned into a word by transposition. In her procedure the stimulus remained in view until the response was executed, so the transposition effect cannot be assigned to short-term memory in any very reasonable way. Therefore discrimination of the order of letters influences the perceptual recognition of printed words. Mason (1980; Note 6) reports that reading comprehension scores are more strongly related to saying where in a brief display a letter occurred than to saying what letter it was. Thus there is evidence to suggest that reading ability and speed or efficiency at resolving letter location and order during word recognition are associated. This information processing &ill fa& within the *The role of information gained from the periphery is not yet completely understcod, but it does appear that some rather gross information, such as word length, is obtained from farther to the right than 10 character spaces. Such Information has been shown by Rayner to influence recognition of words during the next fixation, ln which the charactersthat had been peripheralbecome foveal. It may be, then, that spatial order is almost never complete& unconfounded from temporal order, except perhaps in the first fixadon on a new line of print.

86

Thoms H. Carr

of skills held to be important by the visual discrimination theories of reading failure. E&es has uncovered a partial solution to the problem of resolving letter order. He found that transposition errors were less frequent in orthographically regular strings than in random strings and concluded that the word superiority effect found in a variety of perceptual tasks stems from the use of knowledge about acceptable spelling patterns to help with lhe processing of letter order. Subsequent tachistoscopic studies have obtained an advantage for wrrds and orthographically regular pseudowords when subjects had only to say which of a small, prespecified set of target letters had occurred anywhere in a letter string (Car-r,Lehmkuhle, Kottas, Astor-Stetson, and Arnold, 1976; Spector and Purcell, 1977). This shows that orthographic structure facilitates letter identification as well as letter location. If orthographic structure can aid in the discrimination of both item and order information, one should ask whether reading ability is related to knowledge of orthographic structure and skill at using it. In answering this question, discussion will move from primary visual processes to a consideration of secondary feedback mechanisms, and then to translation from visual to phonological information. In the next section, then, we will move from the domain of the visual discrimination theories to the domain of the recoding theories. domairz

Is orthographic knuwledge related to reading ability?

Knowfedge of orthographic structure comes in two forms, and their possible effects ought to be considered separately. First, there is knowledge of orthography as a visual structure, the ilcceptable spelling patterns that can be seen as combinations of letter shapes on the page, Second, there is knowledge of spelling-tosound correspondence, the set of mapping rules for pronouncing the spelling patterns.j Not all the words of the English language conform to these two kinds of orthographic regularity (though there are languages, such as Finnish, in which conformity is nearly perfect). In English, at least, a substantial ;ilnount of ungeneralizsble knowledge about specific words is 3There is considerable controversy over how these “mapping rules” are represented in the mind and what kind of mechanism makes use of them. Some theorists assume that people know at some level what amounts to an ordered list of rules that are independent of the 4readyastablished internal lexicon. Others, notably Brooks and Miller (1979) and Glushko (19791, maintain that people have instead a set of algorithms or heuristics for pronouncing a new string of letters by analogy to simihrly spelled words accessed in the lexicon. Whatever kind of mechanism embodies the genera&able spelling-tomund relations of a Lnguage, it appears to be separate from the internallexiconitself. Performanceon words and unfamiliarpseudowords can be manipulated independently in tachistoscopic recognition under conditions in which subjects have been shown to rely primarily on phonological recodings to support their responses (Carr,Davidson and Hawkins, 1978).

Building theories of rending abihty

87

required to supplement the structural knowledge that can be applied to large numbers of strings. It appears that both skill at using visual regularity and skill at using spelling-to-sound regularity vary with reading ability, while the use of word-specific knowledge does not. A series of experiments using the physical same-different matching task with simultaneously presented ietter strings, which has been shown to depend upon visual stimulus information rather than phonological or semantic recodings (Baron, 1975; Barron and Henderson, 1977; Pollatsek, Well, and Schindler, 1975), indicates that orthographic regularity facilitates word procening in the visual system (Carr, Posner, Pollatsek, and Snyder, 1979; Carr, Brown, Myers and Koons, Note 7). The right experiments have not been done to say for sure whether poorer readers benefit as much as better readers from orthographic structure in visual tasks. Results from tachistoscopic reports and from visual starch, however, are suggestive. Frederiksen (Note 1) found that good readers were more accurate than pcor readers at reporting orthographically accepta,ble bigrams from a briefly presented array. The difference was larger for low-probability acceptable bigrams than for bigrams that were both acceptable and quite likely to occur. Jackson and McClelland (1979), on the other hand, found no relation between reading ability and reporting random letters. Mason (1975) obtained a similar patEEm in the ability to determine whether a target letter was present in strings that varied in single-letter positional frequency, which is one component of orthographic regularity. Good _-rl- were helped when either the targelt or the distracters occurred at positions in the string at which they were lik.ely to occur in an English word of the same length. Poor readers showed neither benefit. Mason’sdata indicate that differences between good and poor readers in visual search time are smallest when the letter strings are most discrepcant from English spelling patterns dnd increase as structural predictability increases. Massaro and Taylor (Note 2) have recently replicated Mason’s Wndings concerning single letter positional frequency and extended them to another component of orthographic structure, constraints on letter sequencing as defined by Venezky (1970a). A problem with these studies is that both tachistoscopic report and visual search are susceptible to influence by events outside of or subsequent to visual encoding, and the extent to which nonvisual processing actually affects the results of these two tasks has not been calculated. Though phonological recoding or short-term memory could conceivably be the sourse of the relation between reading ability and tachistoscopic report, Maan’s visual search findings are not easily explained in terms of recoding or memory differentials. They may be due, as Mason argues, to differences among readers in visual code formation. If not, they are most likely due to differences in

88

Thomas H. Carr

looking strategies that are controlled by orthographic constraints on the target’s probable location. In either case, it would be an aspect of O~Ographic knowledge and its use to facilitate the accrual of visual information that is reliably related to reading ability in. Mason’s experiment and its extension by Massaro and Taylor. The evidence on spelling-to-sound regularity is more easily interpreted. A multitude of studies have shown that poorer readers are slower and less accurate at pronouncing unfamiliar but orthographically regular pseudowords (mard, dake, leb, throp) but differ Pittle from better readers at pronouncing very familiar real words (e.g., Calfee, Venezky, and Chapman, 1969; Firth, 1972; $;dSOii, 1978; Perfetti and Hogaboam, 1975; Frederiksen, Note 1). Thus poor readers are in good shape when well-practiced knowledge about specific words can be used to gain a pronunciation, busi at a disadvantage when gc:~eralizable or rule governed knowledge must be brought to bear. As me&oned in describing the single-factor theories, some investigators attribu;c poor readers’ problems with phonological recoding to a short-term memory deficit. For example, Jorm (1979) argues that the child “must be able to remember the phonemes which result from applying each graphemephoneme correspondence rule and the order in which the phonemes are to be output. He must also remember which letters of the word he has already analyzed so that he will avoid going over 1ettGrs which have already been accounted for, or omitting strings of letters which have not.” (p. 25). It does seem plausible that short-term memory limitations could play a role in determining success at rule-governed spelling-to-sound translation. But for several reasons remembering the letters and the individual phonemes produced from the letters is probably the least of the poos reader’s worries. There is considerable evidence that the concept of what a phoneme is gives children great difficulty, and that isolating phonemes from the speech signal is very hard (Gleitman and Rozin, 1977; Rozin and Gleitman, 1977). Kindergartners and poor first and second grade readers often have trouble deciding whether a word contains a particular sound or even whether two words rhyme, and the ability to manipulate and make judgments about the sound of letters and words is still correlated with comprehension-weighted reading speed among college students (Jackson and McClelland, 1979).4 Thus, 4Jackson and McCJelhrnd(1979) argue that the basic information processing tasks that they found to account for independent variance in readingability share the need to activate higher-ordermemory codes for perceptual input quickly and accurately. From this they conclude that the speed of accessing higher-order cods+ especially phonological codes, is the major factor distinguishing among readersof varying ability. However, Jackson and McClelland’stasks require manipulation and comparison of phonological codes after they have been activated, and we have already seen that poor readersdo not differ much from good readers in the time and accuracy needed to pronounce very famihar words for (continued on facing page)

Building :I tories of reading ability

09

readers, especially young beginning readers, differ substantially in “phonemic awareness”, and differences in phonemic awareness correlate with rt;ading achievement (Golinkoff, 1978). Becoming phonemically aware, however, does not solve the problem of spelling-to-sound translation. Rules for mapping between phonemes and letters are extremely complex. The vast majority are conditional, depending on such things as the position of the letter in a string, other surrounding: letters, the presence or absence of a silent signal letter, and even the derivational history or usual syntactic function that the string containing the letter miP&t have (Gibson and Levin, 1975; Gleitman and Rozin, 1977; Venezky, 197Oa, 6). Conditional rules are notorioudy hard in concept %rmation tasks anti there are large individual differences among children in how rapidly conditional rules are learned (Bruner, Goodnow, and Austin, 195 6 ; Stevenson, 1972). Therefore mastering and applying the rules of spelling-to-sound translation are tough tasks (Barron, 1979; Doehring, 1976; Gibson and Levin, 1975). What’s worse, the rules alone will not yield enough completely accurate pronunciations for a reader to get by, so children must learn to coordinate the rules with word-specific knowledge (Glushko, 1979; Venezky, 1970a) All of this means that getting phonemes and phoneme clusters to begin with could pose a much bigger problem than remembering them once they have been obtained. Finally, putting phonemes together into a pronunciation is difficult for children and is also associaisd with reading skill (Golinkoff, 1978). Chall, Roswell, and Blumenthal (1963) found that the ability to blend three phonemes into a syllable correlates as highly as 0.6 with reading achievement during the first four grades of school. Blending errors among poor readers often involve a failure to completely drop the schwa from consonants when attempting to combine them with vowels, even when the sequence of consonant/schwa and vowel sounds is produced in the right order (Gleitman and Rozin, 1977; Venezky, 1972). The limits on blending seem to be set by the ability to combine phonemes that cannot be pronounced correctly in isolation, not by the ability to keep their order straight (Gleitman and Rozin, 1977). One has to conclude from this discussion that remembering the raw materials is only a small part of spelling-to-sound translation. Several comwhich a well-learned, directly activatable phonological code already existsin memory. Thus Jackson and McClelland’sdata seem more consistent at present with the hypothesis that poor readerslack phonological analysis skills than the hypothesis that they suffer from slow access to stored higher order codes. It is possible that some of the evidence that has been taken to indicate a deficiency in code activation is confounded by differences among readers in how well-learned are the directly activable codes. Degree of learning will often vary with reading ability unless stimuli are carefully selected, because good readers spend more time reading than poor readers.

90

i%ommH. Curr

ponents of the process other than short-term memory have been shown to be reliably related to reading ability, as claimed by the recoding theorists.

What about memory for the temporal order of words rather than phonemes?

Here is one place where differences in short-term memory could be very important to differences in reading ability. Kleiman (1975) and Levy (1978) have demonstrated that engaging in activities that tie up the speech production apparatus disrupts sentence comprehension, even though Kleiman working with ad&s and Barron and Baron (1977) working with first graders found that judgments about the meanings of individual familiar words were not disrupted by concurrent speech activity. Levy (1978) summarizes the available data by saying that these activities interfere with phonological recoding and short-tenm memory (see also Perfetti, Note 8). While phonological codes are not always needed for access to word meanings, they do prove useful as a memory medium to support comprehension requjring accurate maintenance of word order or close attention to full information content. Since the evidence cited earlier shows that good ,and poor readers often differ in their ability to maintain the order of auditory or phonological sequences, detailed comprehension may suffer among poor readers because of a specific problem with short-term memory for word order. Other arguments along these lines have been laid out by Baddeley (1979) and Ferfetti and Lesgold (1977). The conclusion is supported by Caramazza and Zurif’s (1976) work with brain damaged patients who show large short-term memory deticits but reasonabiy good long-term storage. Caramazza and Zurif found that these patients were quite capable of comprehending contrasts between nonreversible, active and passive sentences, where there is really only one sensible reading given the component parts of the sentence, but failed when the sentences were reversible. Reversibility makes strict maintenance of word order necessary for accurate understanding. A similar deficit was found among deaf readers by Scholes, Cohen, and Brumfreld (1978). These investigators agreed with Conrad’s argument that the absence of phonological coding in short-term memory was responsible for the problem. Though the above position on phonological coding is sensible, it is based largely on differences between good and poor readen; observed in several kinds of short-term memory tasks that often bear little similarity to reading. One might wonder whether these memory differences would actually occur in the reading of connected text. Ferfetti and Lesgold (1977) have adapted Waugh and N9l;man’s (1965) probe recall technique to this question. At various points in the mum of reading a passage, tird and fourth grade

Building theories of read&g ability

91

subjects were interrupted by the presentation of a probe word that had ‘occurred either 4 or 7 words earlier in the passage. The subject’s job was to recall the worci that had followed immediately after the probe word in the text. If the probe occuiyed 4 words back in the text, then 3 words intervened between target and time of test. If the probe occurred 7 wcrds back, then 6 words intervened between target and time of test. This procedure does not completely separate order memory from item memory, but it places a definite premium on maintaining order. Resillts showed that when subjects read silently, good readers enjoyed a 4.5% advantage over poor readers with 3 intervening words and an 18.0% advantage with 6 intervening words. Thus poor readers appear to forget the wording of sentences faster than do good readers in this probe recall task with connected text as the stimulus material. Of course, the probe recall technique requires a type of very complex memory performance that people never have to do while reading outside the laboratory. One could legitimately ask whether some simpler memory judgment, such as forced.=choice recognition, might not be a more appropriate test. Though alleviating some of the complexity of the memory measurement task itself-which could magnify differences among readers beyond their normal proportions-- recognition would seem to move further away from the actual memory demand+ of ongoing comprehension than recall. When one forgets the beginning of an involved sentence or paragraph by the time one reaches the end, it is a failure of recall. The forgotten material may or may not seem familiar when one I~ Mns to the beginning and rereads the passage, but that judbnent is essentially irrelevant to the memory failure that interrupted comprehension and caused the regression in the first place. Thus tradeoffs must continually be made between unwanted perturbations of the reading process introduced by complex measurement techniques and unwanted misrepresentations caused by the choice of some other technique that unfortunately measures the wrong thing. This is not intended either to defend or to attack Perfetti and Lesgold’s use of probe recall-rather, their experiment provides a good opportunity to bring up a pervasive difficulty that faces all researchers in the area. However, it ought to be noted at this point that the size of the poor readers’ disadvantage in Perfetti and Lesgold’s data may be exaggerated by the complexity of the procedures used to measure it. At least two different ways of reducing or eliminating poor Weaders’ deficits in short-term memory performance have been reported in the literature, and looking at them1 briefly will hc!p to illustrate more clearly the nature of the poor reader’s Imemory problem. Torgeson and Goldman (1977’) gave good and poor second-grade readers a task in which the experimenter poiated one-by-one to each of seven line drawings of common objects

on a page of a booklet, waited 15 seconds with the child’s showed the same seven objects arranged in a new random on the next page of the booklet. The child was asked to point to thr in the same order as the experimenter-a test of the child’s memor; poral order. The good readers were considerably better than the poor at this task, but the difference was almost eliminated when the childasked to name each object aloud as the experimenter pointed to it. and Goldman argued that naming the objects aloud enhanced the verbal rehearsal during the 15-second interval, and that the ader advantage occurred because of a metamnemonic differstrategy. The good readers were more likely to spontaneously e absence of the requirement to label than were the poor readers. up this argument, Torgeson and Goldman reported that observabl:: ofrehearsal varied in exactly the same way as reproduction accuracy ups in both task conditions. n and Goldman’s argument may be true, but a metamnemonic difference zems unlikely to explain Perfetti and Lesgold’s abilitymemory for text, This is because it seems unlikely that text they have already read while they are in the text. Yet Perfetti and Lesgold report a very similar in their p_&e recall task. When subjects read the text aloud rather ntly, memory was generally better for everyone, and the advantage readers over poor readers was reduced from 4.5% to 3.5% with 3 and from 18% to 9.5% with 6 intervening words. for Perfetti and Lesgold’s findings might be found ir.:the good and poor readers will engage in phonological 3 r?ber than the likelihood that phonological codes once they have been activated. Generally speatig, lists of onologically similar items are harder to maintain accurateiy ry task than lists of acoustically or phonologically dis1964; Crowder, 1976). Shankweiler and Lieberman etrimental effects of phonological similarity were good readers. The difference held whether stimuli visually or auditorily. Mark, Shankweiler, Lieberman, and reported that false alarm rates to distracters that rhymed in a recognition memory test for words were also lower for readers. Stimuli in this memory test were read by the Shea (1979) have recently extended the finding to runauditorily presented words: poor readers made fewer g distracters than good readers. Despite the paucity of f&e alarms among poor readers, they nevertheless

Building theories of reading ability

93

made more semantically-based false alarms (to synonyms of target words) than did good readers. When the stimuli were changed to auditorily presented pseudowords, phonological false alarms by poor readers increased, but good readers remained more susceptible than the poor readers to phonologicallybased confusions. These data suggest that poor readers are less likely than good readers to employ phonological coding in situations where memory is important. Since phonological encoding is commonly believed to be short-term memory’s best hedge against loss of order information, the poor reader3 difficulty with the maintenance of order may be at least in part a rather direct consequence of deficits in phonological recoding, as Perfetti and Lesgold have argued in their hysteresis hypothesis. However, both Byrne and Shea’s and Shankweiler and Lieberman’s poor readers showed less reliance on phonological codes even when the codes were provided for them via auditory stimulus presentation. Thus the problem does not lie completely in skill at spelling-to-sound translation. Poor readers appear to be less sensitive to phonological codes or less facile with them in general than good readers. Byrne and Shea’s comparison of words with pseudowords indicates that the problem is exacerbated when semantic codes are available that could possibly distract the poor reader’s attention away from the phonological codes that he or she is already having trouble in using. But can deficiencies in the utilization of phonological codes be the entire explanation of the comprehension difficulties that remain after problems with visual code formation have been taken into account? Two comprehension-related phenomena that bear a much less obvious relation to memory maintenance are the ability to identify syntactic relations among words and the ability to use syntactic and semantic context as an aid to recognizing words currently under perceptual scrutiny. There are experiments to indicate that both of these skills vary with reading ability. Discussion of these experiments will shift the focus from data-driven processing to a consideration of the kinds of processes emphasized by the topdown deficit theories. Syntactic knowledge and use of context In Guthrie’s (1973) study of comprehension, poor readers were less able than good readers to choose a word from a set of three alternatives that would appropriately fYl a blank in an incomplete sentence. This occurred even though the two groups were equally proficient according to the GatesMcGinitie Vocabulary Test at recognizing the words from which the sentence frames were composed. HOW should Guthrie’s finding be interpreted? One possibility has already been considered, that poor readers may have had more

difficulty than good reader&in remembering thy words long enough to integrate their meanings into a composite understanding Of the sentence. s%lath,

short-term memory hypothesis could be advanced, and this has been done (JOT, 1979). However, there ase some other po rle’s experiment, Four types of btnks had to be filled requiring either a noun, a modifier, a verb, or a function word. For each blank, the three alternatives were a completely appropriate word, 8 semantically incongruous word from the same form class, and a word from ++different form class whose meaning was related to or generally consistent with the gist of the sentence. If poor readers failed in this task primarily because of short-term memory limitations, then the root difficulty would be a tendency to scramble word order rather than to forget words entirely, as indicated in previous sections. This would interfere with the assessmznt of syntax, but the meanings of individual words would still be active together to give a global impression of semantic cantent. Short-term memory failure, then, should lead poor readers to choose the syntactically inappropriate but semantically related distractor more often than the good readers. However, there was no difference in the relative likelihoods of syntactic semantic errors-just a main effect of reading ability. As Guthrie s steed, this implies that the poor readers were trying to use the same kinds of syntactic and semantic processing strategies as the g”sod readers, but using them less well. Therefore poor readers may not be as sensitive in general to syntactic and semantic constraints on text, rather than failing to respond m particular to information based on word order as a short-term memory hypothesis would predict. If this notion were correct, then good and poor readers mi more on measures of the ability to use context as an aid in processing current and upcoming text than on measures of memory for text they have already processed. Using Sach’s (1967) recognition paradigm, Lovett (197% has compared good and poor readers’ detection of lexical, syntactic, and semantic changes in sentences that were successfully read aIoud in connected discourse. She found no differences between the groups. WhiIc this may me&n that her materials were too easy to tax her subjects sufficiently, it may also mean that relatively short-term memory for laentance content does not always discriminate among readers who vary in skiIl, This conclusion is not necessarily at odds with the idea that reading comprehension may ait times be limited by the ability to maintain exact word order, since Levy’s work indicated that only certain kinds of comprehension and memory tasks place strict enough demands on short-term memory for quch differences among readers to show up.

a

Buildhag theories of reading ability

95

k for evidence that the use of syntactic or semantic context is related to reading ability. Mackworth readers -%th a series of sentence frames, either a noun or a verb. The task was to complete words printed above the sentence on the atives were nouns grouped to one Fide of e were verbs grouped to the uther aide. ed that poor readers searched less selectively rtion of time looking at words in that the differ&ce in eye movement patterns occurred aders took longer to tell the nouns and the verbs apart, they knew what they were looking for as well as the good would convert Ma&worth’s result into more evidence for a ition instead of syntactic knowledge. Cromer readers who had no ohviou- deficit in of comprehension when the words in the sentences were grouped according to their syntactic phrase structure rather than presented with the usual even spacing between words. The same poor readers displayed equivalent comprehension relative to normal presentation when sentences were given word by word on a scroll. In contrast, phrase strilcture grouping had no effect on good readers but word by word presentation hurt them. Cramer’s data show that some poor readers have a tendency that good readers do not have to treat sentences as stdngs of independent words, and that this tendency results from a failure to identify and exploit syntactic relations. Therefore reading ability seems to be related to differences in the spontaneous application of syntactic knowledge. This conclusion is complemented by Graesser, I-Ioffman, and Clark’s (1980) finding that sentence reading times were influenced much more among poor readers than ders by decreases in syntactic predictability, where yredictalatsd from the form class transition probabilities de&red in Stevens and I?umelh ‘s (1975) Augmented Transition Network grammar. Syntactic processing in general may give poor readers trouble. 8emantCknowledge and use

of

context

What about semantic context? Most of the data on this yuestion come from primed lexical decision and pronunciation. In these tasks some kind of context consisting of words or sentences precedes a word to which the subject has to respond. This context is either meaningfully related to the imperative WCXIand therefore appropriate, or unrelated and therefore inappropriate.

96

Thornas

I-3, Cam

A difference in response time for the two conditions demanstrtites sensitivity to meaningful relations between the context and the imperative word. This differc:nce can be called the total semantic facilitation. Sometimes a third condition is included in which the imperative word i? preceded by a semantically neutral warning signal. When this is done the total facilitation can be divided into the benefit that results from an appropriate context and the cost that results from an inappropriate context. Following Posner and Snyder (1975) and McLean and Shulman (1978), benefit may arise entirely from automatic spreading activation while cost indicates the involvement of an attended expectation. With this terminology established, let us turn to the data. At least one experiment has obtained the same amount of total facilitation for good and poor readers (Merrill, Sperber, and McCauley, in press). However, the more common result is that poor readers show greater total facilitation (e.g., Schvaneveldt, Ackerman, and Semlear, 1977; Wet+ 2nd Stanovich, 1978). This is usually taken to mean that poor readers are st least as sensitive to semantic context as better readers and perhaps more s *lsiti,ve. A couple of recent investigations, though, show that the picture ‘oeco~lies L more complicated when costs and benefits are calculated. Becker (Note 9) reports that people performing primed lexical c’ecisldns can be categorized: some enjoy substantial benefit from appropriate c,ontexts and suffer very little cost from inappropriate contexts, while othe IS Elkffer cost without showing much benefit. When tested for reading speed, people in the benefit category are relatively unaffected by increases in passage difficulty. People in the cost category ;ue slowed considerably. Susceptibility to disruption by passage difficulty is not necessarily the same thing as Teading ability measured on a comprehension test, but the two probably have quite a bit in common. Becker’s findings suggest that good readers gain fr. n meaningful context through automatic prooesses while poor readers use context in a more attention demanding, predictive i&ion. Perfetti, Goldman, and Hogaboam (! 979) report some data with which this speculation can be tested. Perfetti and his colleagues found that words were pronounced faster at the ends of sentences telling a story than at the ends of random unrelated lists, and poor readers showed a bigger difference between the two context conditions than good readers. Next a cloze procedure was added to the story version of the pronu;lciation task. Just before each imperative word was presented the subject guessed what the word would be. Guesses were scored as correct, incorrect but appropriate, or inappropriate. Before considering the pronunciation latencies, the circumstances should be established under which each kind of guess would pr&ably be made. An

Building theories of reading ability

97

inappropriate guess is evidence that the context was not comprehended well enough to understand what kind of word was required. An appropriate guess is evidence that the context was comprehended and constraints on acceptable words realized. Finally, a correct guess is evidence that these constraints were so strong that the set of acceptable words was narrowed to a very small number. Table 1 gives pronunciation times for the actual imperative word as a function of the kind of guess made about the word. Times are expressed both as absolute latencies and as percentages of the latency when the guess was inappropriate. The latter measure corrects for the large overall difference in pronunciation times between the good and poor readers. It is clear that good readers were helped ahnost as much by a context from which they could make an appropriate guess as by a context flom which they could make a fully correct guess. Poor readers, though, were only helped substanti4ly by contexts fram which they could guess exactly what word was going to be presented. Thus good readers do appear to utilize context automatically with little involvement of attention. As long as the context is sufficient to put the good readers into the general semantic hailpark, a lot of facilitation will resanlt. Poor readers do not seem to get much of this automatic support from context, but rely instead on attended computations that help if they are right but hurt i’ thev are wrong in any way. Reading ability, then, ;iyspears to be related to variation along a dimension of cognitive flexibility. The dimension ranges from the automatic utilization of very general constraints on acceptability to reliance on single-valued predictions that lead the reader down a costly garden path if they do not work out.

Treble1.

Prsnunciation latencies for good and poor fifth grade readers as a function of L*lozep:rj?mnance EnPerfetti, Goldman, and Hogaboam 3 (I 9 79) experiment. Type of Guess Made in Cloze Task Correct

Incorrect but Appropriate

Inappropriate

--_ Good Readers Absolute Latency Percent of Latency after Inappropriate Guess Pour Readers Absolute Latency Percent of Latency after Inappropriate Guess

824 msec

656 mstx

706 msec

79.6%

85.7%

100%

792 msec

972 msec

1OCclmsec

79.2%

97.2%

:oO$o

98

Thomas H. Carr

Con tex t-sensittvc encoding

SO far the discussion of semantic context has been limited to predictin upcoming words from what has already been read. Semantic relationshi can also have a substantial influence after initial word reco are well underway. For example, Tulving and Thompson (19’43), Barclay, Brasford, Franks, hTcCarrel1,and Nitsch (19941, and Anderson and &tony (1995) have shown that semantic context can lead to chameleon-like shifts in the meaning of word co:lcepts encoded into memory. The perceived sense of a word such as “shovel” or “lady” will vary from phrase to phrase in which it is used (“the steam shovel” cIersus“the toy shovel” or “the little old lady”’ versus “the first lady of the land”). A particularly extreme kind of context-sensitive encoding or encoding specificity occurs with homographs, in which two or more very different meanings are associated with a single “dog’s bark” or “river bank” versus “First spelling (“tree bark” V~I:VAS National Bank”). Current research suggests that when a homograph is read, all 0:‘ its meanings may become temporarily active (Conrad, 1974; Killion, 1978). A selection prccess then chooses one meaning and suppresses the others, basing its choice on the u priori frequencies with which various meanings are likely to occur in conjunction with the actual semantic demands of the context (Killion, 1978; Schvanevsldt, Meyer and Becker, 1996; Simpson, in press). The selection mechanism determines the sense that is consciously perceived and ultimately the sense that is encode ’ into memory. Such a selection process may underlie all encoding specificity phenomena, though its consequences are most pronounced in the case of homographs. Merrill, Sperber, and McCauley (Note 10) have recently presented evidence that good ana poor readers differ in the efficiency of context-sensitive encoding processes. Fifth graders were shown either a single word or a sentence context, then a target word printed in colored ink. FoBowing a Stroop procedure developed by Warren (1972) and Conrad (1974), the task was to nar se the color of the ink. A single-word context such as “cat” could be followed by one of two related targets such as “fur” or “claw”, or else by an unrelat,ed target. Good and poor readen alike were slower to name the color of both related targets than the color of the unrelated target, The critical manipulation involved the sentence contexts, which ended with the same wor& that were ust d as single-word contexts. Each sentence was constructed to be semantically appropriate to the sense of its last word that was emphasized by cme of the targets, and semantically inappropriate to the sense of its last word that was emphasized by the other target. To carry on with the “cat” example, one sentence context was “The girl touched the cat”, which was appropriate to the target word “fur”, while the other was “The girl

Buildingtheoriesof readingability

99

t the cat”, which was appropriate to the target word “claw’“. Given ‘cot” had primed both ccfur” and ‘%law” when presented in isolation, the of the sentencecontext condition would reflect rather directly on of the context-based selection mechanism. Results showed that for good readers, the semantically appropriate context-target combination produeed interference relative to an unrelated combination, but the inappropriate combination did not. For poor readers, however, both combinations produced interference, just as “cat” had produced interference for both ‘&fur’”and “claw” In the single-wotd context condition. Thus contextsensitive selection of a particular sense of the la& word of the sentence seemed to characterize the performance of the good readers, while unselective activation of all senses seemed to characterize the performance of the poor readers. If comprehension and memory depend on integrating the meanings of constituents at various levels of analysis into a coherent semantic repre sentation of a text, this difference in context-sensitive encoding could figure importantly in individual differences in reading ability as measured ny tests of comprehension and memory. A summary of reading-@&d individual differences in cognition

To put a brief end to a long argument, comprehension performance is reliably related to a large number of information processing skills, many of which have been the basis for single-factor heories of individual differences in reading ability. Among these skills are discrimination of letter location and order during perceptual recognition, use of orthographic regularity as an aid to perceptual encoding and decision making, use of spelling-to-sound regularity in phonological recoding and pronunciation, memory for the order of a sequence of words, spontaneous identification and exploitation of syntactic relations, flexibility in the use of semantic context as a guide to prediction, and specificity in the use of semantic context as a guide to encoding. Were the analyses carried beyond the bounds of single sentences, the list might possibly grow to include such things as facility with anaphoric reference (Fmderiksen, Note 1l), knowledge of the usual structure of stories and other kinds .of text (Baker and Stein, 1978; Graesser, Hoffman, and Clark, 1980), or the propensity to supplement explicit textual information by drawing inferences (Baker and Stein, 1978; Brown, et aZ., 1977; Paris and Lindauer, 1976). Therefore individual differences in reading ability may well be multiply determined-they are unlikely to be traced to differences in any one factor, no matter how crucial that factor seems to be to the readihg process. At least the podbthty of more complexity must be admitted to

100

ThomasH. Can-

conceptions of reading skill than the single-factor theories have done, however regretfully one might give up the many satisfaction9 afforded by a simple explanation. Relative contributions of cognitive skills to reading performance Demonstrating that good and poor readers differ on a large number of cog nitive skills, however, does not finish the job that’s necessary. We still need to establish the relative contributions of individual differences in various skills to individual differences in overall performance. There is not nearly as much solid evidence on the relative importance of skill differences among readers as there is on the existence of these differences. In the course of

looking at some of the most recent evidence, it will become necessary to distinguish two rather different kinds of skills that could each be predictive of reading comprehension: general information processing abilities and specific mental operations as they have been developed within the task of reading itself. Evans [ 1979, (Carr and Evans, in press; Carr and Evans, Note 12)I has col-

lected data from which intercorrelations can be calculated among a number of general measures of information processing, maturity of oral language, and reading comprehension in first graders. She included tests of sequential short-term memory, Piagetian class inclusion operations, pattern analysis and predictive reasoning (Raven’s Colored Progressive Matrices), multiplechoice cloze performance (the “comprehension” subtest of the Stanford Achievement Tests, Primary Level 1), mean length of utterance in conversations with an adult about a number of specified topics, syntactic complexity of utterances in the: same conversations (the Developmental Syntax Score of Lee and Canter, X971), and sophistication at verbal description and storytelling (called “quality of expressive language” and defined in detail in Carr and Evans, in press). Table 2 shows the relationship between these measures and reading comprehension, which was assessed by asking fact, inference, and vocabulary questions about a series of short passages graded for difficulty (primer, first grade, second grade). Shart-term memory, class inclusion, and predictive reasoning all correlated positively with comprehension, and they also correlated positively among themselves. Multiple-choice cloze performance was a much better simple predictor of comprehension than any of the nonreading cognitive measures, capturing 846% of the variance in answering questions about the reading passages with its correlation of +0.92, Correlations between cloze performance and the measures of short-term memory, class inclusion,

Buildmgtheoriesof readingabili:ty

Table 2.

Correlation with Comprehension

101

Simplecorrelationsbetweencognitivemeaswesand readingcomprehension amongfirst graders cl;lss

fnclusion

Short-Term Memory

Progressive Matrices

MultipleChotce Cloze

Mean Length of Utterance

Developmental Syntax Score

Quality of Expressive Language

+0.47*

+0.51*

+o 45’

+0.92*

- 0.02

-0.27

+0.23

*p < 0.05.

and predictive reasoning were all much smaller than this (+0.44, +O.S1, and tO.43, respectively), suggesting that cloze performance would remain substantially corlslated with reading comprehension even if the contributions lrlade to both of these reading measures by the nonreading cognitive tests were extracted. While considering the research by Evans and Carr, we should examine the approach itself as well as the data that were collected. The decision to test for relations between fairly general information processing capacities and reading comprehension rests on the assumption that in order to read, one brings to bear a number of highly transferrable cognitive abilities that presumably underlie performance in many other intellectual activities as well. These abilitiesexist independently of whether or not an individual ever learns to read, though specific practice is necessary to get them to function at their best in the service of reading as opposed to other tasks in which they might be involved. How valid is this approach? The appreciable magnitude of the correlations between reading comprehension and r Dnverbal short-term memory, class inclusion relations, and predictive reasoning lends it credibility, and it is one of only two approaches available if one wants to predict in advance of reading experience how successful a learner is likely to be. The other possibility is to provide the novice with a standardized amount of reading instruction and measure how much learning results. Assessment of teachability or learning potential was first suggested by Vygotsky (1926/ 1962), and has been employed with good results by Siegler (1978) in research on the development of reasoning and by Feuerstein (1979) in evaluation ansi remediation of cognitive delay. However, the much higher correlation observed between comprehension and the more obviously reading-based cloze test indicates that the amount of task-specific learning to be acquired is considerable. Once reading experience has begun, one might therefore want to know how vtious task-specific performances are actually coming along, rather than now well they ought to be coming along given the learner’s general inforrrd &ionprocessing abilities and

102

ThornasH.Carr

learning potential. In order to iind tkat out, one needs to supplement general ability and learning potential assessment with a component skills analysis of

the particular task at hand. The first step in such an analysis is to specify the component skills that make up the task, and the second is to determine which of the specified skills could potentially be involved in determining individual variation in the overall success of the task as a whole. These two steps have been the major topic of the present paper. The third step is to determine !he relative contribution to variation in overall success that is in fact made by each skill or skill group. Two major kinds of strategies could be employ ed in taking this third step, and I will describe an example of each. One strategy is to look as carefully as possible at the fdll act of reading, using skill inventories such as the one offered here to educate attention as to what details of performance might be illuminating. By comparing characteristics of the text to changes in performance measures-for example, reading time, miscues in oral production, or comprehension errors-one can infer by judicious reference to the skill inventory just what aspects of the information processing requirements of reading seem to make the most difference to success. Oral miscue analysis has been developed in detail by Goodman (1973) and put to excellent use by Weber (1968, 1970) and Biemiller (1970). This work is already well-known. As a further illustration of the stratogy, we can look at an analysis of reading time done recently by Graesser, Hoffman, and Clark (1980). Passages were constructed that varied along six dimensions. Three of these were considered to be part of the “microstructure” of the passage, and were defined on individual sentences taken in isolation: the number of words in a sentence, the number of propositions, and the proportion of words whose syntactic form class was predictable. Propositions were identified by Kintsch’s (1974) criteria-and syntactic predictability was calculated using Stevens and Rumelhart’s (1975) Augmented Transition Network parsing system. Three mo,;e dimensions were considered a part of the “macrostructure” of the pas:;age, and were defined on relations between sentences: the number of new argument nouns in a sentence, the familiarity of the topic discussed in the passage, and the narrativity or “story-ness” of the passage. New argument nouns were characters, objects, locations, or concepts appearing for the first time (Chafe, 1972), while familiarity and narrativity were assessed from subjects’ ratings. A widely circulated story with a cast of characters and a plot, such as %YOWWhite and the Seven Dwarves”, would be high in familiqrity and narrativity; an exposition conveying information on a relatively obscure topic, such as “Armadillos”, would be low in both. In two experiments, sentence reading times were collected from college students who read 12 passages sentence by sentence using a self-paced pro-

Building theories of reading abi@

103

cedum:. Mctltiple ’ n andyses showed that together the six text variables predicted the total vtimce in sentence reading time in the first ~x~e~rnent an the second, and that each of the variables except syntactic p~di~tab~ity made a si nt ~dependent contribution in both experiments. Syntactic predictab naily significant in one exveraged aca~ss the two experiments, by narrativity (33.3%) and number ment nouns accounted for another 2.5% of the familiarity I A%, and syntactic predictability outcome concerned the relative contributions and propositional integration to the total time needed ion. Reading the 2.42 words in the average proposition contributed 164 msee per word or 397 msec, while integrating the information from the words into a proposition contributed only 148 msec more. It would appear, then, that the processing of a proposition was hugely taken up by word recognition. The integration process itself seemed to be accomplished very rapidly (or else substantially in parallel with word recognition, implying continuous integration of new information as it became available until the proposition was completed. Graesser et al. seem to believe the former, but the latter is also plausible.) After the combined analyses, the subjects in each experiment were divided into two groups according to reading speed. Mean reading time for an average sentence of 12 words was 3.46 set for the faster group and 5.35 set for the slower group. Though subjects were told they would be tested on the material, no test was actually given, Therefore reading ability was inferred from reading speed. Multiple regression analysts were performed fQr each group separately, and the slopes were compared for each variable. Results showed that fast (“‘good”) and slow (“poor”) readers differed on all three microstructure variables but on none of the macrostructure varittbles. Averaged across experiments, slow readers took 1.7 1 times longer to process a word, 2.22 times long&+rto integrate a proposition, and 6~33 f,tnes longer to deal class relative to a predicted one. The with a syntactic&y unpredicted n cost and benefit of semantic context latter result is much like the find discussed earlier. This su sts that the mechanism by which context is used to predict upcoming wo perates upon syntactic and semantic constraints similarly, and that poor readers may show more cost than good readers in syntactic predictions as well as semantic predictions. To put these findings in perspective, it ought to be noted that $he second eyneriment in Graesser’s study included a manipulation of the readers’ expectations about why they were reading the passages. Half were told they would be given a multiple-choice test on the information in the passages and half were told

104 ThornasH. Carr

they would be asked to write essays about them. This maniguhtioa was effective, producing differences in the contributions to reading time made by all three macrostructure variables. It did not change the contributions OP for different microstructure variables. And despite the fact that read’ s, it was still purposes changed the pattern of macrostructure contribu microstructure variables that distinguished fast from slow readers in the experiment. Looking at the relative importance of the three microstructure variables, ater on the difference in efficiency between good and poor readers was both syntactic prediction and propositional integration than on word processing, but the time to process words was the biggest component of total reading time in absolute terms. Taking that into account, it would appear from the data of Graesser et al. that the largest part of the total reading time difference between fast and slow readers at the college level is due to word processing. That is to say, equating fast and slow readers on word processing efficiency would do more to equate them on total reading time than Ghanging any other single skill. Some investigators have argued that the relative importance of word processing skills is greater among less experienced readers than among more experienced readers (e.g., Calfee, Veneaky, and Chapman, 1969; Gibson and Levin, 1975). If that wer2 true, then Graesser’s findings would lead to the conclusion that the development of efficient word processing is critically important -it is the largest single source of variation between fast and slow college readers and it is an even larger contributor to reading ability a+ lower grade levels. Therefore the next study to be described will sen e two purposes. In addition to illustrating another st.:ategy by which the relative eo&ibutions of component skills could be pursued, it addresses in detail the twopart question of how much is contributed to reading comprehension sy skill at word processing and what factors inCuence the development of that skill. The study was done by Singer and Crouse (in press), who colIected independent measures on a number of reading-related tasks from a large group of sixth graders. Thz tasks were discrimination of letter identity, phonemic awareness and phonetic recoding, vocaFaularyknowledge, cloze performance, and the Gates-McGinitie comprehension test. In addition, they administered the Raven’s Pro,gressiveMatrices, a measure of predictive reasoning. The data were then analyzed using multiple regression. TQs procedure for obtaining measures of predictor variables is quite different from Graesser et ul, ‘s. It is actually an extension of the research approach evidenced in the extensive literature already reviewed concerning differences between good and poor readers on individual cognitive ttisks. Both Graesser et al. and Singer and Crouse wanted to estimate the impoF

Building theories of readingabitity

1%

tmcesfvariolls diffe~n~~~ amon readers in the abilitv to perform mental required by reading. Graesser et ai. chose to do that ius materials that they thought would make several operations easier or harder tting subjects to perform al1 operaether in as natural a way ossibfe, and looking for effects of materials on a summary measure of overdtl performance. Singer and Grouse e same goal by extracting several proposed operations process, embodying each in an independent task where d be observed directly, and asking how strongly the direct f skill at each operation related to a separate direct measurement criterion task. The former approach has the virtue of ecoat it p~se~es to a greater extent the usual task environtal operations of interest must be carried out. The latter approach has the virtue of speerfiability and control, in that it allows the investigator to determine more directly the operation that is being observed and to manipulate more systematically the way in which it must be carried out. Because speeifiability and control are important to understanding how each operation works while ecological validity is important to uriderstanding how the operations fit into the complete performance, these are complementary strategies. They might best be viewed as converging operations (rather than as competitors for the allegiance of researchers, for example). Turning to the results of the investigation, simple correlations with comprehension w +0.41 for the Progressive Matrices, +O.f 6 for letter discrimination, +0.3 r phonemic awareness arld phonetic recoding, +0.76 for vocabulary knowIodge, and +-OS4for cloze performance. When the measures were entered in that order into a multiple regression equation, vocabulary knowledge with a standardized regression coefficient of +0.62 and cloze performansc with a coefficient of +0.17 COII tributes1 significantly to the variance in comprehension scores. Singer and Grouse then asked how the remaining information-proc;?ssing tasks were related to the two direct predictors of comprehension. When cloze performance was taken as a ctiterion variable, significant regression eoeffrcients were obtains8 for vocabulary knowledge (H.39) and for a task that did not directly predict comprehension, Progressive Matrices performance (t0.19). These relations indicate that if a reader knows a lot about the words that might be used and can do well at predictive reasoning, then the reader is likely to do well at the cloze task--a very sensible result. Another regression analysis shofved that, like cloze performance, vocabulary knowledge was also predicted by Progressive Matrices performance (+0.40), and, in addition, by recoding skill (tO.40). This, too, seems to be a sensible result. Much has been made about the role that recoding plays as a vocabulary

106

ThornasH. Carr

builder for the reading process. If a reader can pronounce an unfamiliar string of letters, the word might be recognized by phonological reference to the intern& lexicon even if that word does not yet possess an entry that has become visually-addressable through repeated readings (Baron and Strawson, 1976; Carr, Davidson, and Hawkins, 1978; Carr and Evans, in press; Coltheart, 1978; Gibson and Lcvin, 1975; Rozin and Gleitman, 1977). A second vocabulary builder for the reading process involves guessing a word’s meaning from context, which probably depends in part on the competence at predictive reasoning that is tapped by Progressive Matrices performance. The analyses conducted by Singer and Crouse, then, may have identified a hierarchically organized set of interdependencies among several readingrelated information processing skills that could have substantial developmental and instructional implications for the acquisition of reading competence. A rough sketch of this skill hierarchy is shown in Figure 1. Whether such a model could stand up to empirical evaluation remains to be seen. Certainly some of the relational parameters of the model would change with developmental level and reading experience -for inst:mce, among readers younger than those who provided data to Singer and Crouse, such as the first graders observed by Evans and Carr, predictive reasoning might be relatively less L%portant and recoding relatively more important as predictors of comprehension (Calfee, Venezky, and Chapman, 1969; Doehring, 1976; Wanat, 1974). Among older readers such as the college c+~~ Yrrrdentsstudied by Graesser et al., predictive reasoning might gain in relative importance, perhaps through its task-specific instantiations in utilization of semantic and syntactic context. In addition to the need for adaptation of the model to possible developmental changes, a number of cognitive performances in which readers have been shown to vary are not represented at all in the skill hierarchy of Figure 1 (for example, visual discrimination of letter order, use of orthographic structure during visual code formation, short-term memory for word order, and specificity of semantic encoding). Others are submerged in “skills” that we already know to be complex skill groups --for example, phonemic awareness, knowledge of spelling-to-sound translatiola rules, automaticity of the application of spelling-to-sound translation rules, and phonemic blending are all submerged in “recoding”, while spontaneity and flexibility in the use of syntactic context and flexibility in the use of semantic context are submerged in “cloze performance”. Determining how these unrepresented and submerged skills might fit into the model constitutes a major &aIlenge. Since the model as it stands is limited to individual differences in microstructure processing, another major challenge would be to certify whether or not a microstructure mo se1 is really sufficient to account for individual differences in comprehension cf connected text. There is no guarantee that these chal-

Building theories oj reading ability

Figure 1.

107

A tentative model of the hierarchical relationsamong cognitive skillssupporting reading comprehension derived from the data of Singer and C’rouse (in press). Skills enclosed in ovaIs were direct predict srsof comprehension, while skills enclosed in boxes were predictors of the direct predictors but not of comprehension itself according to multiple regression analyses. Numbers associated with arrows show the strength of prediction as defined by regression weights.

predlclwe

roasonmg

letter

idenhty

dlscrlmlnation

lenges could be met. It seems to me, however, that the class of hierarchical skill models represented by the example in Figure 1 is a good class of models to investigate more fully than we have done to date, and that a good way to pursue them is through the converging application of research strategies like those of Graesser, Hoffman, and Clark (1980) and Singer and Grouse (in press).

References Anderson, R. C., and Grtony, A. (1975) On putting apples into bottles-a problem of polysemy. Cog. &“chol., 7, 167-180. Baddeley, A. 1’. (1979) Working memory and reading. In P. A. Kolers, M. E. Wrolstadand H. Bouma (Eds.), I+ocessing of visiblelunguugsI. New York, Plenum. Baddeley, A. D., and Hitch, G. (1974) Working memory. In G. Bower (Ed.), The psyd Aogy of Iearnlne: rnd nm&vuHon(Vol. 8). New York, Academic Press. Baker, L., and Stein, N. (1978) The development of prose comprehension skills. Champaign, IL: Center for the Study of I&*&g Technical Report 102. Barclay, J., Brantiord, J. D., Franks, J., McCarreB,N., and Nitsch, K. (1974) Comprehension and semantic flexibility. J. verb. Learn. verb. Behav., 23,471-481. Baron, J. (1975) Successive stages in word r?eognition. In P. M. A. Rabbit and S. Doxnic (Eds.), Attentionand performanceV. London, Academic Press. Baron, J. (1979) Spelling and reading by rules. In U. Frith (Ed.), Cognitive processes in spelling. London, Academic Press. Baron, J., and Strawson, C. (1976) Use of orthographic and word*pecific knowledge in reading words aloud. J. exper.PsychoI,HumanPercep.Perform.,2, 386-393.

108

ThomusH. Can-

Barron, R. W. (1979) Some aspects of the development of visual word recognition. In T. G. Wailer and G. E. MacKiwon (Eds.), Reading research: Advancesin theory andpractice (Vol. 2). Nelv York, Academic Press. Barron, R. ‘W.(In press) Wsua , orthographic, and phonological strategies in reading and :;r:!!iig. In U. Frith (Ed.), Cognitiveprocessesin spelling. London, Academic Press. Barron, R. W., and Baron, 1. (197’7) How children get meaning from printed words. Child Devel., 48, 587 -594. Barron, R. W., and Henderson, L. (1977) The effects of lexical and semantic information on samedifferent visual comparison of words. Mem. Cog., 5.566-579. Benton, 4: L. (1962) Dyslexia in relation to form perception and directional sense. ln J. Money (Ed.), Reading disability: aogress and research ueeds in dyslexia. Baltimore, Johns Hopkins Press. Biemiller, 1. J. (1970) The development of the ure of graphic and contextual Inform&ion as children learn to read. Read. Res. 0.. 6. 75-96. Bock, J. K., and Irwin, D. E. (1980) Syntactic effects of information availability in sentence production. J. verb. Learn. verb Behav., 19,467-484. Boder, E. (1971) Developmental dyslexia: Prevailing diagnostic concepts and a new diagnostic approach. In H. R. Myklebust (Ed.), Progress in learningdisabilitiesVol. 2). New Yorrt, ? -UPS! and Stratton. Boder, E. (1973) Developmental dyslexia: A diagnostic approach based on three atypical reading spelling patterns. Devel. Med. ChildNeural., 15, 663-685. Brooks, L., and Miller, A. (1979) A comparison of explicit and implicit knowledge of an alphabr:t. In P. A. Kolers, M. E. Wrolstad and H. Bouma (Eds.), Processingof visiblelanguageI. Ntw ‘fork, Plenum. Brown, A. L., Smiley, S. S., Day, J. D., Townsend, M. A. R., and Lawton, S. C. Intrusion of a thematic idea in children’s comprehension and retention of stories. Child Devel., 48,1454-1466. In A. Reber and D. Scarborough (Eds.), Towardsapsychologyofreading. Hillsdale, NJ., Erlbaum. Bruner, J. S., Goodnow, J., and Austin, G. (1956) A study of thinking. New York, Wiley. Bryan, W. L., and Harter, N. (1899) Studies in the acquisition of a hierarchy of habits. Psychol. Rev., 6.345-375. Byrne, B., and Shea, P. (1979) Semantic and phonetic memory codes in beginning readers.Mem. Cog., 7,333-338. capfee, R. (1977) Assessment of independent reading skills: Basic resemch and practical unplications. In A. Reber and D. Scarborough (Eds.), Towarda psychology of readhzg.Hillsdale, NJ..Erlbaum. Calfee, R., Venezky, R., and Chapman, R. (1969) Pronunciationof syrrtbeticwordswithpredictable and unpredictable letter-soundcorrespondence. Technical Report 71, Wisconsin Research and Jkveltipment Center for Cognitive Learning. Cammazza, A., and Zurif, E. (1976) Dissociation of algorithmic and heuristic processes in language comprehension: Evidence from aphasi . Brain Lang., 3.572-582. Cam, T. H. (1979) Consciousness in models of information processing: Primary memory, executive control, and input regulation. In G. Underwood and R. Stephens (Eds.), Aspects ofconsciotrsness (vol. 1). London, Academic Press. Carr, T. H., Davidson, B. J., and Hawkins, H. L. (1978) Perceptual flexibiity in word recognition: Strategies affect orthographic computation but not lexical access. J. exper. Psychol. Hum. Pemp. Perform., 4,674-698. Cam T. H., Lehmkuhle, S. W., Kottas, B., Astor-Stetson, E. C., and Arnold, D. (1976) Target position and practice 51 the IdentJtication of letters Jn varying contexts: A word superiority effect. Percy. Psychophys., 19,412-416. C% ‘I’. H., Posner, M. I., Pollatsek. A., and Snyder, C. R. R. (1979) Orthography and familiarity effects in word processing. J. exper. Psychol. Gen., 108, 389-414.

Building theories of reading ability

109

Carr, T. H., and Evans, M. A. (In press) Classroom organization and reading ability: Are motivation and skill antagonistic goals? In 9. Edwards (Ed.), Social Psych~iogy of Reading. London, Edward Arnold. Carroll., J. B., and Maxwell, S. E. (1979) Individual differences in cognitive abilities. An. Rev. Fsychol., 30.603-640. Chafe, W. (1972) Discourse structure and human knowledge. In P. Frsedle and J. Carroll (Eds.), Languagecomprehension and the acquisitionof knowledge. Washington, DC., Winston. Chal!, J. S., Roswell, F. G., and Blumenthal, S. H. (1963) Auditory blending ability: A factor in success in beginning to read. The Reading Teacher, I 7, 113 -118. Chambers, S. (1979) Letter and order information in lexical ac:nss. J. verb. Learn. verb Behav , 18, 225-242. Chang, F. R. (1980) Active memory processes in visual sentence comprehension: Clause effects and pronominal reference. Mem. Cog.. 8, 58-64. Chomsky, N. (1970) Phonology and reading. In H. Levin and J. P. Williams (Eds.), Basic studies on reading. New York, Basic Books. Chomsky, N., and Halle, M. (1968) The soundpattern of English. New York, Harper and Row. Clark, H. H., and Clark, E. V. (1977) Psychologyand language.New York, Harcourt Brace Jovanovich. Coltheart, M. (1978) Lexical access in simple reading tasks. In G. Underoood (Ed.), Strategies of informationprocessing. London, Academic Press. Conrad, R. (1964) Acoustic confusions in immediate memory. Brit. J. Psychol.,‘.U, 75-84. Conrad, C. (1974) Context effects in sentence comprehension: A study of the subjective lexicon. Mem. Cog., 2, 130-138. Conrad, R. (1972) Speech and reading. In J. F. Kavanaugh and I. C. Mattmgly (Eds.), Language by ear and by eye. Cambridge, MA, MIT Press. Corkin, S. (19’74) Serial ordering deficits 111inferior readers. Neuropsychol., 12, 347-354. Critchley, M. (1964j Develonmentaldyslexia. London, Heinemann. Critchley, M. (1970) 77zedysZexicchild. London, Heinemann Medical Books Ltd. Cromer, W. (1970) The difference model: A new explanation for some reading difficulties. J. educat. Psycho/!,61.471-483. C&owder, R. G. (1976) Rinc@les oflearning and memory. Hillsdale, NJ., Erlbaum. Crowder, R. 6. (1979) Similarity and order in memory. In G. Bower (Ed.), 77zepsychology o/leurfiing and motivation:Advances in researchand theory Vol. 13). New York, Academic Press. Davidson, B. J. (1978) Coding of individualwords during reading. Unpublished doctoral dissertation, Department of Psychology, University of Oregon. Doehring, D. (1968) Patterns of impairment in specific reading disability. Bloomington, kldiana University Press. Doehring, D. (1976) The acquisition of rapid reading responses. Mono. SIX Res. Child Devel., 41 (2, Serial 165). Estes, W. K. (1975) The locus of inferential and perceptual processes in word identificati0n.J. exPer. Psychol. Gen., 104, 122-145. Evans, M. A. (1979) A comparative study of young children’s classroom activities and learning outcomes. Brip. J. educat. Psychol., 49, 15-26. Feuerstein, R. (1979) ?%e dynamic ossessmentof returded performers. Baltimore, MD, University Park Press. Fir& I. (1972) Components of reading drrubility. Unpublished doctoral dissertation, University of New South Wales. Frith, U. (1979) Reading by eye and writing by ear. In P. A. Kolers, M. E. Wrolstad and H. Bouma (Eds.), Processingof visiblelanguageI. New York, ?!enum. Gibson, E. J., and L&n, H. (1975) The psychology of reuding.Cambridge I%% MIT Press.

Gleitman, L., and Rozin, P. (1977) The structure and acquisition ofreading 1: GrtB structure of language. In A. S. Rebet and D. Scarborough (Eds.), Towa& reading. Hi&dale, NJ, Erlbaum. GPushko, R. J. (1979) The organization and activation of orthographic knowhrdge in rea J. exper. Psychol.Hum. Percep. Perform., 5.674-691. Golinkoff, R. (1978) Phonemic awarenessskills and reading achievement. In F. Murray and J. Pikulsk (Eds.), The acquisitionof reading. Baltimore, University Park Press. Goodman, K. (1969) Reading: A psycholinguislic gub%inggame. In K. Goodman and J. Fleming (Eds.), Selected papers from rhe Z&4 Preconvei ‘ion Ins&We. Newk, BE: International Reading Association. Goodman, K. (1973) The 13th easy way to make learning ‘o read difficult: A reaction to Gleitman and Rozin. Read, Res. Q., 8,484-493. Goodman, K. S. (Ed.) (1973) Miscue anufysis: Application to reading instruction. Detroit, Wayne State University Press. Goodnow, J. J. (19723 Rules, rqertoires, and tricks of the trade. In S. Famhamory (Ed.), Informationprocessing in children. New York, Academic Press. Graesser, A. C., Hoffman, N., and Clark, L. F. (1980) Structural components of reading time. J. verb. Learn. verb. Behav,, 19, 135-191. Guthrie, J. T. (19?3j Reading comprehension and syntactic responses In good and poor readers. J. educat. Psychol., 65, 294-299. Huey, E. B. (1908) l7rc psychology und pedagogy of reading. New York, Macmillan,(Cambridge, MA,

MIT, 1968). Jackson, M. D., and McClelland,J. (1979) Processingdeterminants of reading speed. J. ex@%Psychol. Gen., 208, 151-181. Jan&r, R. (1970) Effects of syntax on running memory span for connected discourse. Psyclron. Sci., IG, 235-236. Jarvelia, R. (1971) Syntactic processing of connected speech. J. verb. Learn. verb. Rehav., JO, 4O9416. Jorm, A. F. (1979) The cognitive and neurological basis of developmental dyslexia: A theoretical framework and review. Cog., 7, 19-34. Katz, L., and Wicklund,D. T. (1972) Letter scanning rate for good and poor readerr in grades two and six. J. educat. Psychol., 63, 363-367. Kiliion, T. H. (1978) Activutlon of multiple meanings in the pmcesslng of lexlculombigulty. Unpublished doctoral dissertation, Department of Psychology, University of Oregon. Kintsch, W. (1974) ?%erepresentationof meuning In memory. Hillsdale,NJ., Erlbaum. Kintsch, W., and van Dijk, T. (1978) Toward a model of text comprehension and production. ps)clrol. Rev., 85, 363-394. Kleiman, G. (1975) Speech recoding in reading.,t verb. Learn. verb. Rehav., 14, 323-339. ck LaBerge, D. (1979) The perception of units in beginning reading. In P. A. Weaverand L. 1** (Eds.), The theory und pmctice of early reading: An IntroductSon (Vol. 3). Hill Erlbaum. Lee, L., and Canter, S. (1971) Developmental sentence scoring: A clinical procedure for e&ir~;~\ng syntactic development in children% spontaneous speech. J. Sp, Hapr. D~tord., 36, 31.5 -341. Levy, B. A. (1978) Speech processing during reading. In A. Lesgold, J. W. Peliegrlno, S. D. FOkk8t.ta, and R. Glaser(Eds.), Cog&&e psychologyand Insmcrim. New York, Plenum. Lovett, M. W. (1979) The selective encoding of sen tential information in normal reading dtwelopmant. Child DeveI.,50, 897-900.

Mackworth, J. F. (1972) Some models of the reading process: Learnezsa and skilled readera.Read. Res. Q., 7,701-733.

Building theories of reading ability

Mark, L SS*~h~k~~~,

111

D.. %jberm~n~ I. Y., and Fowler,6‘.A. (1999) Phonetic recoding and reading cts of orthographuzstructure defined by Psychol.

Cen.,

fa4, 146-166. stun?readersas a function of readerability and two forms Cog., &,568-581.

d the encoding of item and location information. J. exper. spatial order memory and item memory in six& g&dechildrenas a function of Peadintability.J. educat. Psychof., 67,610-616.

and Jackson,hi. (1998) Studyingindividualdifferenmsin reading.In A. LesgolE J. W. o, S. D. Fokkema, and R. Gaser (Eds.), Cognitivepsychology and instruction. New Y ark, Pbnum. MeLea&J. P., and Shubnan,G. L. (1999 On the construction and maintenance of expectancies. Q. J. e~cper.Psychd., SO. 441-4~4. Menill, E., Speruer,R. D., and McCauley,C. (In press)The effects of context on word Identificatio.1 in good and poor readers.J. Psychol. B’Neill, G., and Stanley, G. U. (1996) V&al processingof straight lines in dyslexic and normal children.&it. 1. educat. Res 46.323-329. Orton, E. (1966) Wo&blindness in school children and other papers on strephosymbolia. Pomfret, CT, The&ton Society. Paris,S. G., and Lindauer,B. K. (1999) Construciiveaspectsof children’s Comprehensionand memory. In 6,. B. Kail,and J. W. Hagen(Eds.),Perspectives on rhe development of memory and cognitfon. Hills&la, NJ., Erlbaum. Perfetti,6 A. and Hogaboam,T. (1995) Relationshipbetweensingleword decodingand readingcompre tsnsionskill.J. e&cat. Psychol., 67,461-469. Perfetti,0 ir., and Lesgold,A. (1998) Discoursecomprehensionand sourcesof individualdifferences. In M. Just and P. Carpenter(EL.). Cog&&e processes in comprehension. Hillsdale, NJ., Erlbslrm .

Perfotti,C. A., Goldman,S., and Hogaboam,T. (1999) Readingskill and the identificationof wordsin discoursecontext. Mem, Cog Pollatsek,A., Wall,A. D., and !Mindlor, R. M. (1995) Fmiliarily affects visualprocessingof words. J. cxprr BsyckoL-Hum.pc,Uzp.Pe@om, I, 328 333. Posnor, H. I., and Snyder, C, R. R. (1975) Attention and cognitive control. In R. L. Solso (Ed.), Inform&ion processing and cognition: The Loyoh~ Symposium. Hi&dale,NJ., Erlbaum. Raymor,K. (1978) Eyo movementsin readingand informationprocessing.Psychol. Bul., 85,618-660. Rosniek,L. (1976) C?ICI natureofintelligence. Milledale, NJ., Erlbaum. Roti, P. (1976) The avelution of inte ttce and accessto the cognitiveunconscious.In !#.Sprague and A. N. Epstein (Ed&),&own ia psychobiolot!y, and physiologlcul psychsilow (vol. 6). Now Ronln,

York,

Acu!wnb

Press.

P., and Gleiunzn, L. (1979) The rtrxtuae and acquisitionof readingII: The readingprocess

and the acquit&ionof the alphabeti!:principle. In A. S. Reber and D. Scarborough(Eds.), Tow&o pslycllrclogy of reudhg. HiMala, NJ., Erlbaum. Rozin, P., P&sky, 5.. and Sotsky, R. (1971) Americanchildrenwith readingproblemscan e&Y learnto readEn:$ishreprwnted by Chinesecharacters. Sci., I 7,1264-l 267. Rum-t,. D. E. (197:‘) Under.~turd&and nunmartzingbriefstodes. In D. Lahge and S. J. Samwls (Edr.),&a pwcesses b reading: Peweption and comprehens,on. Hillsdale, NJ., Erlbaum. Rumelhart,D. E. (1978) An interscdvo model of reading.In S. Domic (Ed.), Artenrfonand per&rmuncz VI. Hillw’Ue, NJ., Erlbaum. .>

112

Thomas

H.Cm

pects of muwkd &mu Sachs, J. S. (1967) Recognition memory for syntactic and SCmmtiC Percep. Psychophys.,2,437-442. Scholes, R. J., Cohen, M., and Brumfield, S. (1978) Some possible Cnu~eS of Whlctic defbittr in the congenitally deaf English user. Amer. Ann. DeaX I23,528-535. Schvaneveldt, R., Ackerman, B., and Semlear,T. (1977) The effects of semantic context on chitdren’s word recognition. ChildDevel., 48.612-616. Schvaneveldt, R., Meyer, D., and Becker,C. A. (1976) Lexical ambiguity, semantic context, and visual word recognition. J. exper. Psychol.: Hum, Percep. Perform., 2.243-256. Schankweiler, D., and Liberman, I. Y. (1972) Misreading:A search for causes. In J. P. Kavanaughand 1. G. Mattingley (Eds.), Language by ear and by eye: l?re rekizticonshapsbetween speech and reading. Cambridge,MA.,MIT Press. Schankweiler, D., and Libemlan, I. Y. (1976) Exploring the relations between readingand R. M. Knights and D. J. Bakker (Eds.), The neuropsychology of learning disorders. University Park Press. Siegler, R. S. (1978) The origins of scientific reasoning. In R. S. Siegler (Ed.), Chi1dren’s thin&&g: Whatdevelops? Hillsdale,NJ., Erlbaum.

Simon D. P., and Simon, H. A. (1973) Alternative uses of phonemic information in speIling. Rev. educat. Res.:.43, 115-137. Simon, H. A. (1979) Information processing models of cognition. Ann. Rev. Psychol., 30.363-396. Simpson, G, (In press) Processing lexical ambiguity with and without context. J. verb. Learn. verb. Rehav.

Singer, M. H. (1979) Further commeut on reading, language, and learning. Harvard E&cat. Rev,, 49, 125-128.

Singer, M. H., and Crouse, 1. (In press) The relationship of context-use skills to reading: A case for an alternative experimental logic. ChildDeuel. Smith, F. (1971) Understanding reading. New York, Holt, Rinehart, and Winston. Smith, F. (1973) Psycholinguisticsandreading. New York, Holt, Rinehart, and Wmston. Spector, A., and PurcelI, D. G. (1977) The word superiority effect: A comparison between restricted and unrestricted alternative sets. Percep. Psychophys. 22,323-328 Stanley, G., and Hall, R. (1973) A comparison of dyslexics and normals in recaIlingletter arrays aftor brief presentation. &it. J. educat.Psychol., 43, 301-304. Stevens, A. L., suld Rumelhart, D. E. (1975) Errors in reading: Analysis using an augmented transition network model of grammar. In D. A. Norman and D. E. Rumslhart (Eda,), Bxpfomtions in cognition. San Francisco, Freeman. Stevenson, H. W. (1972) Children’slearning.Englewood Cliffs, NJ., PrenUce4IaB. Stemberg, R. J. (1977) Intelligence, informationprocessing,and analogi& reasoning: Tlracomponan~ tialanalysisof human abifitfes.Hillsdale,NJ., Etlbaum, Tall4 P. (1980) Auditory temporal perception, phonics, and reading disabilitloe in cbiIdren. kin Lang., 9, 182--191. Taylor, S. E. (1965) Eye movements while read ‘“9: Facts and fallacies,Amer. &cut. R@s,J., 2, 187202. Torgeson, J. K. (1978) Performance of readingdisabled children on serial memory tarb: A tesvtow. Reud. Res. Q., 19.57-87.

Torgeson. J. K.,and Go1dman.T. (1977) Rehearsaland short-term memory In readin8 disabled children, Child Devei., 48,56-60. Ttig,

E., and Thompson, D. M. (1973) Encoding specificity and retrieval processes in episodic memory. Psychol.Rev., 80,352-373. VeMino, F. R. (1977) Alternative conceptualizations of dyslexia: Evidence in support of a verbaldeficit hypothesis. HarvardEducat. Rev., 47,334-355.

Buildingtheotiesofreadingability

VellutinQ, F. R‘,

113

ste

3.‘ and Kandel,G. (197%)Readingdisability: An investigation of the pereeptuaid&bit thesis.Cortex,(5‘106-i 18. Vene~ky,R. (19780) 7&l?$Mt~titslofhhgl&R oFthogruphy. The Hague, Mouton. and spethng Pn H. Levin and J. P. Williams(Eds.), Basic ~e~~ky* R. (197QtrI Re8ularity in rea studiesm ?~~d~~~b&wYerk, VYgetSkY, L. S. 41!%% lltraugA? mf

Ibngum?. Cambridge,MA., MIT Press, (origii&y pubiished

1926). Wanat, S. (Ed.) ($974) fnt&gmcc

und rt?&n& New2rk, DE., International Reading .2ssociation. ami mrrmory. J. expr. Bqychol., 94,90-100. Primary memory. Psychal. Rev., 72,89- 104. R. M. (I%81 The study of oral reading errors. A survey OCthe literature. Read. Res. Q., 4,

58. (f972) Slfmdus call2 , N. and Normaniat, D. A. (

96-119. xt in reading. r‘n H. Levin and J. P. Williams (1970) mrst of grammaticdc (Edr.), R&e stud& ng. New York, Bas West, R. F., and Stanovich,K. W. (1978) Automatic contextual faci.litation in readers of three ages. C&i& Dcvel., 49, 717-727. Yule, W., Rutter, hf., Berpr, M., and Thompzxq J. (1974) Over and underachievement in reading: Distribution in the 8eneral popuhrtion. Brit. J. educat. PsychoI.,44, l-12. f* R* M.

Reference Notes 1 Frederiksen,1. (1978) Assesment of perceptual, decoding, and lexical skiUsand their relation to reading prot%ziency. TechnicalReport1, Bolt, Beronek,and Newman, Boston, MA, USA. 2 MawO, D. W., and Taylor, G. A. (1979) Reading ability and the utilization of orthographic strut twB in ftdh& T~chnic@lRepert 515, Wisconsin Research and Development Center for IndividualizedSchool@;. 3 Friedrich,F. (1979) lne study of skiRedreading: Current issues and methods. Unpublished manuscript. Departmentof Psycholofty,University of Kantas. 4 Singer, hf. H., Allen, W. ‘I’.,and Imppin, J. S. (1975) Differential abilities of good and pcor readers in disc&ninat@ spatial and temporal sequences. Southeastern Conference on Linguistics, NpehvUe, : N, USA. 5 Mm&, F., and MorrisonrF. (1979) Identity and position coding in normal and disabled readers.

Psycltsnp nit Society, Phoenix, AZ, USA. 6 Mason, M. t-9791 Readin8ability and early sensory nrocessfng differences. Psychonomic Society, Phoenix, AZ, USA. 7 Cnrr, T, H,, Brow& T. LepMyers, R., turd Koona, D, Orthography anc~ramiliarity revisited: What happens when wordsurehard to rc? Manuscript in preparation. 8 Issrfettf, C. A. (1979) Phonological codes in reading. Psychonomic k~,iety, Phoenix, AZ, USA. s in word recognitionand reaumg.PsychonomicSociety, 9 Becker, C. A. (1979) Semanticrtrr Phoecti, AZ, USA. 10 Merril4,E., Sperbur,R. D., and McCauley,C. Differencesin semanticencoding as a function of roadin8comprehendonskill. Manuscript submitted for publication. 11 Frederiksen,1. (1979) Undexstandinganaphora:Rules used by readers in assigning pronominal mfemnts. Paychonomic So&y. Phoenix, AZ, USA.

114

12 Carr, T. H., and Evans, M. A. Coherence and dissociation among cognitive skius that underlie reading. Manuscript in preparation.

JusquV ce jour la plupart cles thkories sur l’aptitude 1 la lecture n’ont propos6 qu’un seul facteur en tam que source essentielle des differences individuelles. L’accord SW le facteur a et6 faible et parmi les princiraux candidats propos& on trouve la discrimination visuelle, le recodage phonologique et semantique, la memoire $ court terme et l’utilisation du contexte linguistique. Dans cet article on resume les theories i un seul facteur et on passe en revue la Wrature pour montrer qu’aucune thdone ~3un seul fxcteur ne peut dtre adiquate. La rCussite en lecture est correl%e i de multiples savoir-faire. Parmi ceuxci les capacitds de: discriminer la place des lettres et leur ordre pendant la reconnaissance perceptuelle. de n’utihser la r&ularit8 orthographique que comme aide pour la formation d’un code visuel, d’utiliser les r6gularitk entre appelation-sons dans le recodage phonologique, de m6moriser l’arche des mots, d’identifier spontakment les relations syntaxique et de pr6dire de facon simple 1 park du contexte syntaxique et s6mantique la spbificit6 co-rtextuelle dans l’encodage dmantique. On conclut a la nkessit6 d’un mod&e complexe multifactcriel et otnprdsente certains essais ricents pour rasscmbler les donndes contribuant $ un tel mod&, ‘Trois auproches ont itd d&ies pour identifier les facteurs pertinents pour la rdussite dans la lectu:r- Nvaluation des capacit& ginkles, l%volution dcspotentialit& d’apprentissage et l’analyse componenwlle des savoir-faire. Deux methodes d’analyse des savoir-faire sont pr&ent8es et on recommande de les utiliser dans des op6rations convergentes. Enfm on utilise les kultats de l’analyse des savoir-faire pour proposer un premier exemple d’une classe de meddles hiirarchiques de la capacite de lecture qui peut %tre 6tudi6 de facon develop mentale.

Cognition, 9 (1981)115-116

115

Books received

P. Fletcher and M. Garman (Eds.), LunguageAcqutsition,Cambridge, Cambridge University Press, 1979. R. Jakobsonand L. R. Waugh,7?zeSIwndShapeof Language,Harvester Press, Sussex, 1979. H. C. Furth, TFre Worldof Gnwn-Ups. t3iidren’s C&tceptionsof Society, ElsevierNorth Hollanli, Amsterdam, 1979. B. Harrison, An Introductionto the Philosophyof Language,Macmillan Press, London, 1980. H. J. Klausmeier & Assocs., CognitiveLearningandDevelopment:Informationfiocessing and &getian Perspectives.Batlinger Publishing Co., Cambridge, Mass., 1980. A. Streri, Enonciation et RPf!reuciation:Etude Gt!n&iquedes Productionsd’Pnon&s avec ‘%ouZoir”et“dire’: Editions du CNRS, Paris, 1980. M. M. Gruneberg, P. 13.Morris and R. N. Sykes, fiacticalAspectsof Memory,Academic Press, London, 1980. M. Klein, LtvesPeopleLive - A Textbookof Transactional Analysis,Wiley, Sussex, 1980. D. N. Robinson, Syste,msof Modem Aychology: A CriticalSketch, Columbia University Press, New York, 1980. F. J. Varela, Rinciples of Riologi2alAutonomy, North-Holland, Amsterdam, 1980. C. Tanz, Studiesin the Acquisitionof Deictic Terms, Cambridge University Press, Cambridge, 1980. P. Sangstad, A T?aeoryof Languageand Understanding, Universitets Forlaget, Oslo, 1980. P. H, Leiderman, S. R. Tulkin and A. Rosenfeld (Eds.), Cidturein Infancy: Variationsin the Human Isicpetience,Academic Press, New York, 1980. J. 3, Lass (Ed.), Speech and Language:Advancepin BasicResearchandfiacttice,Academic Press, Ney York, 1980. M. Sazzaniga, D. Steen and B. T. Volpe, Bnctional Neuroscience, Harper and Row, New York, 1980, J. Gliedman and W. Roth (For the Carne@e Council on Children), The Unexpectcr; Mtnotity,Harcourt, Braceand Johanovich,NewYork, 1980. F. C. Keil,Semanticand ConceptualDevelopment- An OntologicalPerspective,Harvard University Press, Cambridge, 1980, D. Cooper, The Lunguage of Madness,PenguinBooks,Harmondsworth,Middlesex,1980. P. Vroon, fntelllgence: On Mythsand Measurement,North-Holland, Amsterdam, 1980. P. Watson, Waron the Mi&, PenguinBooks,Harmondsworth,Middlesex,1980. M. Denis, Les fmugesMentales,Presses Universitaires de France, 1980. A. N. Perret-Clermont,SocialIntern&on and C&nit&eDevelopmentin C%iidren,Acadelnic Ress, New York, 1980.

116

J. T. E. Richardson, Mental Imagery and Human Memory, Macmillan Press, London, 1980. M. P. Marcus, A Theory of Syntactic Reco&tion for Natural Language, MIT Press, Cambridge, Mass. 1980. J. Byan and F. Thomas, ?Jze Politicsof the Mental Handicap, Penguin Books, Harmondsworth, 1980. D. C. Dennett, Brainstomts, Bradford Books, Montgomery, Vt., 1980. C. E. Marks, Commturotomy, Consciousness avid Unity of the Mind, Bradford Books, Montgomery, Vt. 1980. E. Walker (Ed.), Explorations in the Bioiogv of Language, Bradford Books, Montgomery, Vt. 198t.X D. Caplan (E&j, BiologicalStudies of Mental Fkocesses,MIT Press, Cambridge, Mass. 1980. F. Affolter and E. Stricker (Eds.), Perceptual Processes as Prereqwisitiesfor Complex Human Behavior, Hans Huber Verlag, Berne, 1980. L. Kardos and Csaba Pleh (Eds.), Problems of the Regulation of Activity, Akemiai Kiado, Budapest, 1980. J. J. Gibson, The Economic _, Approach to Visual Perception, Houghton Mifflin, New York, 1980. W. Bailey, Man, ReZigionand Science, William Baley Publishers, Santa Barbara, California, 1980. L. Y. Fyans (Ed.), Achievement Motivation: Kzent fiends in Theory and Research, Plenum Press, New York, 1980. 0. H. Mowrer (Ed.),pSychology ofLanguageandLearning, PlenumPress,New York, 1980. K. V. Wilson, From Associations to Structure: The &urse of Cognition, North-Holland, Amsterdam, 1980. R. W. Rieber (Ed.), Applied Psycholinguisticsand Mental H&t?z, Plenum Press, New York, 1980.