Journal of Experimental Psychology...

Journal of Experimental PsychologyMonograph

Vol. 80, No. 2, Part 2 May 1969

GROUP STRUCTURE, CODING, AND MEMORY FOR DIGIT SERIES1

GORDON H. BOWER " AND DAVID WINZENZ

Stanford University

Recognition and recall of digit series were studied as a function of segmentalgroupings imposed on the series either by the location of pauses or in thenaming of successive numerical groups, e.g., 1735 was read to S as "seven-teen, thirty-five." Experiments show that alteration of group structure ofthe same underlying digit string severely degraded memorial recognition ofits repetition and that the normal improvement in immediate recall withrepetition was annihilated by changing groupings at each presentation.Although a second presentation of a string with altered groupings is notrecognized as a repetition of its earlier occurrence, this event is equivalentto an exact repetition when they are assessed by 5"'s later ability to. recognizean ungrouped version of the underlying string. Repetition with the samegroupings establishes one strong trace, whereas repetition with changedgroupings establishes two weak traces either of which may mediate recog-nition of the uncoded version of the string. The "reallocation" hypothesiswas proposed as a summary of these results, whereby group structure affectsperceptual coding, which determines "where" the trace of the event is stored.This was contrasted to a "bin" hypothesis for serial recall. Experiments todifferentiate these involved recall of strings in which only a subsequence orportion recurred. As predicted by the reallocation hypothesis, recall of therecurrent constant chunk improved only when it was located at the beginningof the string.

The following experiments concern the coded as "thirty-seven, nine hundred forty-relationship between group structure, per- six."ceptual coding, and recall and recognition Although this segmental coding might bemeasures of memory for a digit series. A considered to reflect the transfer to digitbasic strategy that 5s apparently employ in series of Ss' habits of performing syntacticallearning an arbitrary series of symbols is to parsings of sentences, the division of a digitsegment or group successive items into sub- series into subjective chunks may serve sev-jective chunks, emphasized perhaps by im- eral purposes for the learner. First, theplicit vocal stress or pausing. With digit subjective chunks are small (2-4 digits) andseries, this segmental parsing may even be are themselves easily learned, so 5"s task isreflected in the words 5" uses in representing reduced in part to seriating a smaller num-the series, e.g., the series 37946 might be ber of units than was true before groupings

were imposed. Second, the groupings mighti This research was supported by a grant (MH- be imposed in a manner calculated to pro-

duce a rhythmic p*tern- For examPle- aRequests for reprints should be sent to Gordon series of 10 digits might be so segmented as

H. Bower, Department of Psychology, Stanford to produce a repetitive triplet-doublet pj-it-University, Stanford, California 9430S. tern, such as (xxx) (xx), (xxx) (xx),

1

© 1969 by the American Psychological Association, Inc.

GORDON H. BOWER AND DAVID WINZENZ

where accent marks are placed over vocallystressed symbols and parentheses representphrase-marking pauses.

Neisser (1967) has speculated on thefunctional significance of such rhythmic pat-terns. First, the accented beats and thephrase-marking pauses could serve as dis-tinctive "anchor points" to which individualsymbols are attached, thereby enabling 5" tokeep track of their relative location in theseries. Second, the overall rhythmic pat-tern might serve as a hierarchical plan forgenerating the serial list (cf. Johnson, 1965;Yngve, 1960). That is, the sequence con-ceivably could be generated by recursive de-coding into, e.g., first vs. last constituentsalong the lines suggested by Yngve (1960)and Miller and Chomsky (1963) for sen-tence generation.

Regardless of the current credibility ofthese speculations, it is sensible that thebasic units which are assembled by a serialrecall mechanism are going to be subjectivegroups. This more limited view was testedin the following experiments. One obviousimplication is that successive symbols withina subjective group will be more stronglyassociated than successive symbols that goacross group boundaries. Available evidencesuggests this is true. 'For example, Muellerand Schumann (1894) showed strong posi-tive transfer when S learned a new seriesderived by piecing together functional groupspreviously learned in different series, butlittle transfer when the new series was de-rived by using adjacent elements from dif-ferent groups of the previously learned series.Norman 8 found, with the probe-digit recalltechnique for series, that associations betweenElements n and n + 1 were high if both ele-ments occurred within a group (of stressedpairs), but were low if the adjacent pairoccurred across group boundaries. McLeanand Gregg (1967) reported on free emissionof a grouped series, finding that interresponsetimes were short in recalling successive ele-ments within a group, but were longer be-tween group boundaries. Such results sug-gest that perceptual groupings imposed by5" or E define the boundaries of functionally

SD. Norman, personal communication, 1967.

integrated response units. In addition, Gar-rett (1965) found that a pause in an audi-tory digit series determined "perceptualunits" insofar as a superimposed click wasperceived as having occurred closer to thepause than it in fact did (cf. Fodor & Bever,1965).

If it is true that S learns a series by im-posing stable functional groupings on it, itfollows that learning could be seriously re-tarded by forcing 5 to adopt different group-ings each time the same series recurs. It isthis prediction which was tested in the firstthree experiments: For some series the samegroup structure recurred repeatedly, whereasfor other series the group structure differedevery time the series recurred. The predic-tion was simply that repetition of a serieswith the same group structure would leadto improvements in its immediate recall,whereas repetition with a changing groupstructure would produce relatively little im-provement in recall.

These studies required that E largely con-trol the groupings adopted by S. A varietyof methods are available for achieving thisend: For visual presentation, E could intro-duce spaces, dashes, or parentheses to markthe groups, as in 17-586; with auditory pre-sentations, E could say the digit names withrhythmical stress and pausing, as in "oneSEVEN . . . five eight SIX," or £ could saythe common names of the numerical groupswith pauses, as in "seventeen . . . five hun-dred eighty-six." Intuitively, the lattermethod, saying the group names to 5", ap-pears to afford the strongest control over 6"sgroupings, so this method was adopted in allbut one of the experiments. This directphonemic coding of groups should producethe largest behavioral effects; the alternativemethods may be expected to produce similareffects to the extent that they determine 5"sgrouping of the series.

EXPERIMENTS I AND IIThe first two experiments investigated the

benefit for recall of repeating a digit serieseither with the same or with a constantlychanging group structure. Experiment Iused 9-digit strings; Exp. II used 12-digitstrings in order to produce a larger effect

GROUPING, CODING, AND MEMORY FOR DIGITS

of the variable. The design was one intro-duced by Hebb (1961) and since studied byMelton (1963) and others, in which a par-ticular digit string periodically recurs amonga set of constantly changing items. Interestcenters on improvement in immediate recallof the recurring string.

Method

Experiment I.—Ninety permutations of the digits1-9 were taken from the tables of Moses and Oak-ford (1963) disallowing series with more than oneadjacent pair in natural serial order. Each nine-digit string was subdivided haphazardly into suc-cessive digit groups of Sizes 1, 2, 3, or 4; theonly restriction was that the number of groups ineach string not exceed five, but within this re-striction there was considerable variation in thenumber of groups, the number of groups of eachsize, and the sequence of group sizes in each nine-digit string. These 90 items were assigned to 10blocks of 9 items; one of these items was designatedto be the recurrent item and was presented onTrials 3, 6, 9, and 12 of the block, and the re-maining 8 items were designated as noise items,presented once each. With letters representingdistinct digit strings, each block followed the para-digm abCdeCfgChiC, where C is the recurringitem in the block. For 5 of the 10 blocks, the re-current item retained the same groupings over itsfour trials. For the remaining 5 blocks, the re-current item had its group structure changed overits four trials. For example, the recurrent itemcoded as (17) (683) (945) (2) on its first occurrencemight be coded as follows on its next three pre-sentations : (176) (8) (394) (52), then (1) (768)(39) (452), and then (1768) (3) (94) (5) (2). Thesegrouped digit strings were recorded on a Wollen-sak tape recorder by saying the numerical nameof each group followed by a brief pause. Timingrelations are difficult to specify exactly, but allstrings were read in approximately 6 ± .5 sec.

Individual -S"s received instructions about immedi-ate recall and the blocked nature of the experiment,but no mention was made of recurrent items. Afterthe warning signal "Ready," they were to listento each string, and when a terminal click oc-curred immediately after the last number, theywere to begin writing the individual digits in left-to-right order in nine blank spaces provided on ananswer sheet. The 5 was told that each seriesconsisted of a permutation of the exhaustive set1-9. He was further told that his recall scorewould be the number of digits written in theircorrect location and that he could either guess ator leave blank those positions he could not remem-ber. He did not have to indicate the groupingsheard, just the underlying digit sequence. Therecall period was 7 sec. (usually more than ampletime) ; there was no feedback, and 5" covered uphis successive recall lines with a cardboard beforehe heard and recalled the next series.

The 10 blocks of items, 5 with repetition sameand 5 with repetition different, were given to5s in different random orders. There was a 30-sec. pause between blocks while E advanced thetape recorder to the starting place for the nextblock.

The 5s were 12 undergraduate students fulfillinga service requirement for their introductory psy-chology course.

Experiment II.—The design was similar in allrespects to Exp. I with the following changes:(a) Each series was 12 digits long; and (6) theblock size was eight strings long, symbolizedaBcBdBeB where the recurrent string B appearedfour times separated by a noise string. The 12digits in a string were obtained from the first 12entries in tables of random permutations of thenumerals 1-20 (Moses & Oakford, 1963), deleting10 and 20 and rewriting 11-19 as 1-9. Thus, eachdigit type from 1 to 9 could appear zero, one, ortwo times in the string, and the number of digittypes with two tokens varied from three to fiveover the strings. Two consecutive tokens of thesame digit were disallowed. Each string of 12digits was divided into no more than seven groupsof Sizes 1, 2, 3, or 4 in a completely haphazardfashion. The grouped digit series were tape-recorded at the approximate rate of one string in7.5 sec. There were 10 blocks of trials conformingto the aBcBdBeB paradigm. In 5 of these blocksthe recurrent string was repeated with the samegroup structure (RS), and in 5 it changed (RC).Recall time was 10 sec. and the interblock restintervals were 30 sec. The 5s were 12 under-graduate students from the same source as Exp. I.

ResultsExperiment I.—Recall protocols were

scored in terms of two indices: (a) whetheror not the entire string of nine digits wasrecalled correctly and (b) the number ofdigits recalled in their correct absolute posi-tion. There was relatively little change overthe 10 blocks of the experiment, so the blockswere pooled. Using the notation abCdeCfg-ChiC to denote each block, there are threeitem types: (a) the once-presented noiseitems (N), distinguished as the first two(ab), the second two (de), etc.; (£) Item Cwhen it recurs with the same group structureover its four trials, abbreviated RS for "re-peated same"; and (c) Item C when it re-curs with a different group structure on eachof its four trials, abbreviated RC for "re-peated changed."

Figure 1 shows the results of the firstrecall measure, the proportion of strings thatare recalled perfectly over the four trials


..30

.20

• Sam* Grouping ̂ x/ix

.Changed Grouping

1 2 3 4Trials

FIG. 1. Proportion of nine-digit strings totallycorrect in immediate recall over four presentations.

within each block. The points for the Rcurves are each based on 60 (12 Ss X 5blocks) observations, while the points forthe N curves are each based on 240 observa-tions. For the N items, there was a sig-nificant decline over the four portions of theblock, F (3, 33) =3.98, p < .05, possiblyreflecting within-block proactive interferencethat dissipates with the rest interval betweenblocks. The feature of interest in Fig. 1is a marked improvement over trials in recallof the recurrent item repeated with the samegroup structure and a nonmonotonic curvefor the recurrent item repeated with differ-ent groupings. The Trials X RS vs. RCItems interaction is significant, F (3, 33) =6.53, p < .01, indicating a greater learningeffect for the RS items. Average propor-tions of totally correct recalls over the fourtrials are .44 for RS items, .31 for RC items,and .33 for N items. The RS items exceedthe latter two, but the RC and N items donot differ significantly.

A similar picture emerges if recall is scoredin terms of the number of errors per string.Error scores averaged around 1.95 out of9 possible. Combining the first two vs. lasttwo trials, the respective mean error scoreswere 2.09 vs. 1.13 for the RS items, 1.94 vs.1.90 for the RC items, and 1.80 vs. 2.05for the N items. The "learning effect," in-dexed by a reduction in error scores fromthe first to the second half of the trials, is.96, .04, and -.25 for the RS, RC, and Nitems, respectively. Pairwise t tests on theseerror reduction scores showed the RS items

to be significantly different from the others,if (11) =4.53 for RS vs. RC, but the RCand N items did not differ significantly,t (11) = 1.13, p> .20.

The results of Exp. I confirmed expecta-tions. A series repeated with the same groupstructure was recalled progressively betterdespite the intervention of two noise itemsbetween presentations. A series repeatedwith differing group structures showed littleimprovement, being recalled no better thanonce-presented control items. This happeneddespite the fact that -S" was writing downexactly the same sequence of digits eachtime the recurrent item occurred (albeitwith different input groupings).

Experiment II.—The principal results areshown in Fig. 2 in terms of the mean errorsin recall per 12-digit string. As expected,the base level of errors on N items wasquite high, around six per series. As inExp. I, there was a significant within-blockregression in recall of the N items, F (3, 33)= 4.20, p < .05. The RS and RC items dis-play recall trends similar to those in Exp. I;there is a sharp reduction in errors overtrials for the RS items, but virtually constantrecall over trials for the RC items. The RSvs. RC X Trials interaction is highly signifi-cant, F (3, 33) = 8.00, p < .01, indicatingthe differential learning in the two cases.Figure 2 shows that Exp. II succeeded in its

DifferentGrouping

FIG. 2. Mean number of errors per 12-digit stringin immediate recall over four presentations.


main purpose, viz., demonstrating a dramaticimprovement for RS items and virtually nonefor RC items from a low-base-line perform-ance.

Further analyses were done to locate thespecific source of improvement over trialsfor the RS items. To this end, a measureof adjacent item association was used, viz.,the conditional probability that Element n +1 was not recalled correctly given that Ele-ment n of the series was recalled correctly.This transition error probability (TEP) wascomputed separately for all n to n + 1 transi-tions which fell within a group (calledwithin-group TEP) and for all those n ton + 1 transitions which went across groupboundaries (between-group TEP). The leftpanel of Fig. 3 shows these two measuresfor the four N items of each block. Thewithin-group TEPs are uniformly lowerthan the between-group TEPs, indicatingstronger adjacent associations within thanbetween groups. Over the four N items ofeach block, the within-group TEPs remainedapproximately constant, whereas the be-tween-group TEPs exhibited a steady rise,revealed in the increasing recall errors forN items in Fig. 2. The RC items yieldedfunctions similar to the N items, but thesewere more irregular and are not presented.A graph of these statistics for the RS itemsappears in the right-hand panel of Fig. 3.Again, the within-group TEPs began lowand remained relatively constant over thefour trials; the between-group TEPs beganhigher, but progressively decreased to verynear the level of the within-group TEP.

N ITEMS

Between Grouping*

Within Groupings

RS ITEMS

_.Between Grouping*

•J

Within Grouping*

2 3Trial*

2 3 4Trials

100

jI 76

)

[so

I; 25'

FIG. 3. Conditional probability of an error onDigit n H-1 given correct on Digit n for 12-digitstrings over four presentations.

I 2 3 4 S 6 7 6 9 10 II 12Serial Petition

FIG. 4. Percentage of digits recalled at eachserial position for noise items grouped in a 23232pattern.

These data indicate that the main benefit ofrepetition for RS items was in improvingthe transitions between successive groups,but that repetition helped relatively little inintegrating the elements within each group.By analogy to textual material, the digitgroups appear to be acting much like words,and the effect of repetition is to build upcompounds or higher order strings of words.

Serial position curves.—It is of interestto investigate how the serial position errorcurve is affected by the grouping of theseries. The recall of those noise stringswhich had the same group structure waspooled. A serial position curve for oneexample is shown in Fig. 4, where the groupsizes were 2, 3, 2, 3, 2. Other group struc-tures displayed similar patterns. Althoughthe general bowed shape of the classic serialposition curve is apparent, there are alsoclear bumps and troughs in the curve cor-responding to the location of the imposedgroupings. A more revealing analysis ofthe serial recall is the TEP statistics shownin Fig. 5. The TEP plotted over Index iis the conditional probability of an error onElement i given a correct recall of Elementi — 1 of the series. At i = 1 is plotted sim-ply the error probability of the first digitin the series. The transitions moving intoa new group in the series (the between-groupTEPs) are crosshatched. It is clear thaterrors predominate in recalling across groupboundaries and that transition error prob-abilities decline over successive elements


.73

?60

: ASSij .30!

: .15

0- ni& I B 9 10 Jj 12

Serial Position

FIG. 5. Conditional probability of an error onDigit n +1 given correct on Digit n for itemsin Fig. 4.

within the same group. The TEP profileis in fact very similar to those reported byJohnson (1965) for sentence recall, wherehis constituent phrase boundaries correspondto the present digit groupings. Even in thisTEP analysis, however, there is still a de-cided serial position effect on between-groupTEPs, i.e., the crosshatched bars increase,then decrease over serial position.

There were insufficient observations onparticular group structures for RS items toanalyze in detail which between-group TEPsshowed the larger improvements with repeti-tions. A plot of serial position error curves,averaged over all RS items, showed markedimprovement over repetitions in recall prob-ability at all serial positions except the lasttwo (which had very high recall probabilitiesevery trial).

Summary

In summary, the first two experiments haveshown devastating effects on recall of alteringthe phrase structure of digit strings. The nor-mal improvement in recall with repetition waspractically annihilated by changing the groupstructure at each repetition. The results sug-gest a critical relationship between perceptualcoding of material and the learning effect nor-mally induced by its repetition; viz., for multi-trial learning to occur, the material must becoded in substantially similar ways over suc-cessive experiences with it. The results sug-gest a hypothesis which supposes that the per-ceptual coding of the input material determinesa metaphorical "location in memory" at which

it is stored. If two serial sequences are codedin substantially similar ways, the second inputis shunted to the same storage location, thereto make contact with and strengthen the traceof the first input of this series. The hypothesisrequires the further assumption that immediaterecall of a series in this type of experiment ismediated by the strength of the trace in themost recently activated storage location. Thus,a series that is coded and stored in a consistentmanner can accumulate trace strength to im-prove recall, whereas a newly coded RC itemis shunted to a new location, and its recall fromthat location is similar to that of a once-pre-sented noise item.

For convenience in the following, this will bedubbed the "reallocation" hypothesis. It is pri-marily a heuristic and a clearly metaphoricalmodel, envisioning separate storage locationsto which input series are shunted by a codingprocess, which experiences leave behind tracesat the storage location; and a recall mechanismthat generates a response using the informationin the trace at the most recently activatedstorage location. Moreover, earlier discussionand the TEP analysis suggest that the tracestored in a given location is in fact a complexbundle (hierarchy?) corresponding to the indi-vidual groups of the structured series. A lessmetaphorical formulation would suppose thatthe present method for imposing groupings infact sets up a direct correspondence between agrouped series and a sequence of inputphonemes. This view would then suppose thatit is this phoneme sequence which is rehearsedand stored as a chain of associations. When anold digit series is regrouped, it now producesa new phoneme string; this fails to "contact"the trace of the phoneme string used previously,and hence no repetition effect is observed in itsrecall. This phoneme view of matters is at-tractive because other evidence (cf. Adams,1967) suggests that immediate memory is af-fected by phonemic variables. However, Exp.Ill introduces some complications to this simplephoneme view of matters.

EXPERIMENT III

A relevant question is whether the nulllearning observed with RC items in Exp. Iand II depends critically on the directphonemic coding of the number groups inthe repeated sequence. According to thephoneme view discussed above, the digitgroupings were relevant only insofar as theydetermined the sequence of phonemes which

GROUPING, CODING, AND MEMORY FOR DIGITS 7

•c 7^_ A

§ 5

UJ

I2 .S '

Different Grouping

^XL--*-—ip6r--«—~-(

Noise Items

Same Grouping

2 3Trials

FIG. 6. Mean number of errors per 12-digit stringin immediate recall over four presentations.

5 heard. Experiment III therefore com-pared learning of RS and RC items with an-other grouping method that does not directlyvary the phonemic structure of the inputstring. This alternate method was simplyto read the digit names with pauses betweengroups. Thus the sequence 185-62, formerlyread to 51 as "one hundred eighty-five, sixty-two" in Exp. I and II, was read to 5" as"one eight five . . . six two" in Exp. III. Inthis case, for RC items 6" heard exactly thesame sequence of phonemes — only the loca-tions of the pauses changed over repetitionsof the same series.

MethodProcedure. — The design, procedure, and mate-

rials were identical to those in Exp. II, with only

I

100

78

90

" I 2 3 4 5 6 7 8 9 10 I I 12Serial Position

FIG. 7. Percentage of digits recalled at eachserial position for noise items grouped in a 23232pattern.

the grouping procedure changed. All 12-digitstrings were rerecorded by saying the digit names(approximately 3/sec rate) with a distinct pause(approximately 1 sec.) between groups. Durationof the strings was 8-10 sec. depending on thenumber of groups (and pauses) in the string. The5s were 12 undergraduates from the same sourceas Exp. I. Their instructions were identical tothose for Exp. I and II.

Results

The mean recall errors per string areshown in Fig. 6 for the three conditions, N,RC, and RS. The overall pattern is strik-ingly similar to that in Exp. II (cf. Fig. 2).Errors on RS items decrease, on N itemsincrease slightly, and on RC items remainrelatively constant over trials. The RC vs.

.75

i.60

45

.30

.15

flI 2 3 4 5 6 7 8 9 10 II 12

Serial Position

FIG. 8. Conditional probability of an error onDigit n + 1 given correct on Digit n for items inFig. 7.

RS Items X Trials interaction is significant,F (3, 33) = 11.01, p < .01, replicating thedifferential learning rates observed in Exp.II for these two cases. Further analyses ofTEP statistics for N items showed serialposition patterns remarkably similar to thosein Exp. II. One example pattern, 23232,yielded the serial position curve and TEPsshown in Fig. 7 and 8. This pattern maybe compared to Fig. 4 and 5 for the samegroup structure but with phonemic coding ofthe groups. The overall similarity of thetwo patterns suggests that simple pauses in-duce functional groupings just as stronglyas does phonemic coding.

It may be concluded that the null learningof RC items is not a simple consequence ofdirect alteration of the phoneme sequence

8 GORDON H. BOWER AND DAVID WINZENZ

produced when different groupings dictatedifferent numerical names for the groups be-cause the same null effect is produced whenRC items involve the same input sequenceof phonemes, differing only in the locationof pauses. Thus, the temporal segmentationof the sequence determines the recall units,and these units shift as the segmentationshifts.

A defensible view is that although theinput string is digit names with pauses, thesegmentation comes to be represented interms of 5"s implicit naming of numericalgroups. Thus, "one seven . . . " is renamed"seventeen" by 51. This is plausible; itwould also salvage the phoneme theory sinceit would be claimed that 5s in Exp. Ill werestimulating themselves in the manner that Ewas stimulating 5 in Exp. II where RCitems showed no benefits from repetition. Aminor problem with this account, however,is 5V introspective reports in Exp. Ill; mostof the 6"s claimed that they were not doingthis phonemic coding into names of numericalgroups. But such introspections shouldprobably be interpreted with caution.

In any event, with this hedge about im-plicit phonemization, it now becomes difficultto discriminate between the phoneme hypoth-esis and the reallocation hypothesis. In thefollowing, they will be treated as roughlyequivalent, but results will be discussed interms of the reallocation concepts.

EXPERIMENT IV

Supposing the correctness of the realloca-tion hypothesis, what does it imply aboutrecognition memory? The simplest viewwould be that recognition of identity de-pends on the new experience being shuntedto a storage location where there is an oldtrace. If the old trace there is strong enoughto effect a substantial match to the newinput (e.g., the matching score exceeds astatistical decision criterion), then 5 willreport that he recognizes the current input asa substantial repetition of a series he ex-perienced earlier. If the group structureand coding of the input string determine thestorage location to which it is shunted, itthen follows that recognition memory foridentity should be seriously degraded by

altering the input coding between the firstand the second presentation of the sameseries. This implication was tested inExp. IV.

This account places recognition of identityin a pivotal role with regard to the beneficialeffect on recall of multiple repetitions. Afailure to recognize the current input as "old"provides a fairly clear indication that itsrecall will be divorced from the potentialbenefits of prior presentations. Some evi-dence is available with respect to this impli-cation. First, and anecdotally, i"s in Exp.I, II, and III never spontaneously reportedrecognition of identity for repetitions of RCitems, but often did for RS items. Second,experiments by Martin (1967) and Bern-bach (1967) have shown for paired-associatelearning that when 5" fails to recognize thestimulus term as "old," his paired-associateresponse on that trial is a sheer guess. More-over, the subsequent course of learning forsuch items is best characterized by sayingthat the occurrence of a recognition failure"resets" the item back to the beginning,where S started with total ignorance of thatitem. Third, experiments by Kintsch andMorris (1965) and Kintsch (1966) on mul-titrial recognition learning have shown thatrecognition failures act like recurrent "re-setting" events over the trial series; once arecognition failure occurs, practically theentire preceding history of reinforcementswith that item can be written off as virtuallyineffective for later recognition responses.

Experiment IV used the continuous rec-ognition paradigm introduced by Shepardand Teghtsoonian (1961), in which 6" givesrecognition responses to a long series ofitems, half of which are repeats of earlierones. Half of the items were RS and halfRC repeats, and each type was tested atlags ranging from 1 to 24 intervening itemsto provide full information about the reten-tion curve.

MethodDesign and procedure.—The .? heard a series of

305 grouped five-place numbers, had to judge eachas new or old, and assign one of three confidenceratings to his judgment (guess, moderate, certain).The five-place numbers used were the first fiveentries in tables of random permutations of the


Same Grouping

I 3 6 9 12 24Number of Intervening Items

FIG. 9. Probability of a five-digit string beingrecognized as old on its second presentation.

digits 1-9. No 2 items were permitted to havemore than the same three digits in identical serialorder, e.g., the pair 71284 and 12853 would be dis-allowed. There were 155 items in all; ISO werepresented twice, and 5 filler items were pre-sented once, accounting for the 305 in the series.Each item was subdivided into two, three, or fourgroups at random, and the sequence of numericalnames of these groups was recorded on tape;reading was at a rate of approximately 3 sec. foreach five-digit string. For 75 of the repeateditems, the second presentation of the item wasread with the same group structure (RS) ; foranother 75, the second presentation involved amarked change in group structure (RC). Thetwo item types were distributed randomly through-out the input series. The trial sequence was soarranged that approximately one-sixth of eachtype of item had its second presentation at lagsof 1, 3, 6, 9, 12, or 24 intervening items after itsfirst presentation.

The 16 5"s were undergraduates from the samesource as Exp. I. They were assigned in randomalternation to two groups of 8 ,?s. After standardinformation about the task, the 8 Ss in the "naive"condition were told that the strings would be readgrouped, but that they should make their recogni-tion ratings with respect to the underlying digitseries irrespective of its grouping, The 8 5s inthe "informed" condition were further forewarnedexplicitly that for half of the items the secondpresentation would be grouped differently than thefirst presentation, and they should try to concen-trate only on the underlying digit string in makingtheir recognition judgments.

Results

The main results are shown in Fig. 9giving proportion of recognition ("old") re-sponses to RS and RC items at Lags 1-24,averaged over the two instructional condi-

tions. The average false-alarm rate on newitems was .42, so performance on RC itemsstill exceeded chance even at Lag 24. Rec-ognition of RS items obviously exceededthat of RC items, as predicted. The RS vs.RC difference appears to affect recognitionin nearly a constant manner independent oflag; i.e., the probability curves are approxi-mately parallel over lags. This parallelismholds as well when performance scores aretransformed into the d'a of signal detectiontheory (cf. Wickelgren & Norman, 1966),which represents an alternative measure ofrecognition performance.

The overall results in Fig. 9 are in thedirection predicted by the reallocation hy-pothesis. The declining curve for RS itemspresumably reflects the declining strength ofthe trace of the first presentation, whereasthe RC curve reflects in addition a loweredprobability that the second presentationwould be shunted to the same storage loca-tion as the first presentation.

From the viewpoint of the reallocationhypothesis, the puzzle in these data is whyrecognition of the RC items is as good asit is. If the second presentation of an RCitem is coded differently from its first pre-sentation, why is it not treated as a newitem, with a "recognition" probability atonly the chance false-alarm rate ? A hint asto a possible answer to this question is pro-vided by analyzing the data for the informedvs. naive Sa, who differed in degree of in-structional preparedness for the RC items.

.9

.8

1.7

.4

Changed Grouping-Informedv ^

False Alarm Rate "-»,,._*M— • i.— .—.•-**.. »>•—.— >*~»*

Changed Grouping-Naive..«......

I 243 6 9 12Number of Intervening Items

FIG. 10. Probability of five-digit string beingrecognized as old on its second presentation.


Recognition performance for these two sub-groups on RS and RC items is displayed inFig. 10. Within each group there is a largeand nearly constant separation between RSand RC items, but the informed group isapparently performing at a somewhat higherlevel than the naive group, particularly on theRC items. The false-alarm rate was similarfor the two groups: .43 for informed 5"s and.41 for naive 5"s. An overall analysis of vari-ance on hit rate (arc-sine transforms) in-corporating instructional set, RS/RC, andlag, however, yielded significant effects atthe .05 level for only the RS/RC and lagvariables. Instructional set had no signifi-cant main effect, nor did it enter into sig-nificant interactions with the other variables.

Recognition of RC items may perhaps beunderstood in terms of a receding strategywhich several 5"s reported using. Thisstrategy was simply to recede £'s groupedfive-digit string into a standard format nam-ing the individual digits. Thus, the inputstring "eighty-nine, three hundred forty-one"would be implicitly receded and rehearsedas "eight, nine, three, four, one." And, ofcourse, the RC phrase "eight, nine hundredthirty-four, one" would be mapped into thesame phonemic string under this standardreceding. To the extent that this decodingstrategy is applied, it would tend to nullifythe effects of the RS/RC variable and toproduce recognition of RC items. If in-formed 6"s used this strategy more often thannaive 5s, then they would differ mainly onrecognition of RC items. Although theirresults differed in this direction, it was nota significant effect statistically.

This suggested factor could be checked outin further experiments specifically designedto better control 5"s coding strategy. Theplausibility of such strategic complications,however, should not be permitted to obscurethe main result of Exp. IV, viz., that alteredgrouping of the digit string had a drasticallydegrading effect on recognition of identityfrom memory. In reference to Exp. II, thisrecognition of repetition for RC items wasprobably much less than in Exp. IV due tothe greater length of the strings (12 vs. 5digits) and the instructional set of theearlier -S"s.

EXPERIMENT VA further interesting implication of the

reallocation hypothesis was tested in Exp. V.It concerns recognition of a visually uncodedtest string after 5 has experienced two audi-tory inputs of the string, with the groupstructure of the two inputs being the same(RS) or different (RC). For example, afterauditory input of (89) (341), and then (8)(934) (1), 5 would be tested for recognitionof the visual string 89341. Suppose that Swill recognize this visual string as an "old"series that he has heard if he can match itsuccessfully to a trace he has in memory ;assume further that his probability of obtain-ing a successful match for such a string isan increasing function of the number of dif-ferent locations at which this series has beenstored and the strength of the trace at therelevant locations. That is, 5 will recognizethe RC item 89341 if he can match it eitherto the trace of (89) (341) or to the trace of(8)(934)(1). Presumably, he not onlycould recognize it then, but could also tellus something about the way the string hadbeen grouped.

By this account, comparison of recognitionon such uncoded tests for RS and RC itemsthen amounts to comparing 5's ability to findone "strong" trace vs. one of two "weak"traces. In the absence of more detailed in-formation about the processes, one cannotpredict in advance which of these cases willyield better uncoded recognition, or indeedwhether they will be equivalent. However,one determinant prediction is that recognitionmemory in both these cases should exceedthat for an item presented only once. In theterms used above, this once-presented itemwill be represented by a weak trace in onlyone location, and it should be less likely tobe found and matched than would be astrong trace in one location or a weak tracein one of two locations. Experiment V wasundertaken to test this prediction.

MethodDesign and procedure.—Each 6" received 10

blocks of input—test phases, the blocks separatedby 30-sec. rest intervals. In the input phase ofeach block, S listened to 12 auditorily groupedstrings (group names), each string five digits long.The test phase which immediately followed con-


sisted of 12 slides, each visually presenting 1 five-digit string for recognition with reference to theimmediately preceding phase of auditory inputs inthe block.

The five-digit strings and their groupings weresimilar to those used in Exp. IV. Tape recordingof the 12 strings for the input phase consisted ofsaying the grouped string (3 sec.), pausing for3 sec., and then saying the next grouped string.The 12 auditory inputs consisted of the following:(a) the first and last buffer items which werepresented once, but were never tested; (b) twoitems presented once and tested; (c) two itemseach presented twice with the same groupings(RS) and tested; and (rf) two items presentedtwice but with different groupings on the repeti-tion (RC) and tested. The number of items inter-vening between repetitions of the same string wasthree, four, or five, averaging out to four. Theaforementioned items account for 6 of the 12recognition test slides. The remaining 6 testslides were new five-digit distractors that hadnot been previously presented in the experiment.The order of input and test items was so arrangedthat the number of intervening events (inputs ortests) between the last input of an item type andits test averaged out to 9, with a triangular dis-tribution over lags of 9, ±1, ±2, and ±3. Thespecific digit series instancing the RS and RCconditions were completely counterbalanced over^s, i.e., a given string was an RS item for 8 .9sand an RC item for the other 8 6~s. The 10 blocksof items were presented in the same order to all 6"s.

The 5s were 16 students from the same sourceas Exp. I and were run individually. They wereinstructed to decide whether or not each visuallypresented five-place number represented the sameunderlying digit string as some one of the auditorilypresented strings in the preceding input phase.The 5 responded "Old" or "New" and added oneof three confidence ratings (guess, moderate, posi-tive), but received no feedback regarding the accu-racy of his judgments.

Results

Over the 10 blocks of the experiment,recognition performance improved signifi-cantly; there was about a 9% decline infalse-alarm rate and a 13% rise in correcthits. Despite these trends, however, the rankordering of the item types remained the sameover the course of the experiment. Theaverage proportion of old responses to thefour item types are shown in Table 1, pooledover the 10 blocks of the experiment.

Pairwise t tests on the proportions inTable 1 reveal the following pattern: (a)Items presented once were recognized moreoften than were distractor items, p < .01;(b) items presented twice were recognized

TABLE 1RECOGNITION PROPORTIONS FOR UNCODED ITEMS

Item type

Noise items (false alarms)One-presented itemsTwice- presented items

Repeated sameRepeated changed

Proportion

.40

.57

.70

.74

more often than items presented once, bothp's < .01; but (c) recognition of the RS andRC items did not differ significantly, t (15)- 1.42, p> .15.

The results confirm the prediction thatuncoded recognition of RS and of RC itemswould exceed the once-presented items.Also, in this instance, presentation of thesame grouping twice was approximatelyequivalent to presentation of two differentgroupings once each. Indexing the learningeffect by the improvement in hit rate over thefalse-alarm rate, the results fall into a simplequantitative pattern: one presentation of astring produced an increment of .17 over thefalse-alarm rate, while two presentations ofthe string produced an average of about twicethe increment, or .32.

Exp. IV and V together represent a curi-ous reversal of effects. In Exp. IV, changingthe coding of a series markedly degraded itsrecognition, whereas in Exp. V, repetition ofa series with altered coding produced justas good recognition memory as did repeti-tion with the same coding. The difference,of course, is in the nature of the recognitiontest, whether the test string is coded differ-ently than the memorial string which it mustcontact or whether the test string is uncoded.

Although generally consistent with theseresults, the reallocation hypothesis does notparticularly clarify the processing compo-nents engaged when S recognizes an uncodedvisual series as identical to a coded auditoryseries he heard earlier. One possibility isthat 5 tries out several (random) articula-tory codings or parsings of the uncoded teststring, attempting to match these trial codesto some phoneme sequence he has stored inmemory. There are several speculative elab-orations possible on this theme. An alterna-tive possibility is that 5 tries to recede every-


thing during input and test into the samestandard format, viz., the series of names ofthe individual digits. Then the auditorystring "eighty-nine, three hundred forty-one"would map onto the same phoneme sequenceas does the visual string 89341. Moreover,if 5"s were to note increasingly the utility ofthis strategy as the experiment progressed,one could explain the observed improvementin recognition performance over the session.A problem with this account, however, isthat unless the decoding strategy is con-sistently applied, it would predict poorerrecognition for RC than for RS items. Inthe data, however, RC items were recognizedslightly better than RS items, and this rankordering was consistent over the 10 blocksof the experiment. A further implication ofa "decoding strategy" explanation is thatwhen S recognizes a visually uncoded string,he should be unable to recall anything aboutthe phonemic grouping of the auditory inputto which that string refers. The alternateexplanation, i.e., generating trial articulatorycodings for matching tests, would imply that.S" would have this grouping informationavailable when he recognizes an uncodedstring. (Obvious "corrections" for guess-ing or false alarms would have to be con-sidered.) Thus, future experiments mightyield more differentiating evidence if 5 wererequired also to indicate the grouping thatthe test string had had during input.

EXPERIMENTS VI AND VIIExperiments VI and VII return to the

immediate recall paradigm of Exp. II in anattempt to better elucidate the mechanismunderlying the repetition effect in serial re-call. Experiments I, II, and III indicateda practice effect for RS items but not forRC items. The reallocation hypothesis ex-plained the RC results by supposing that adifferently coded series is assigned to a dif-ferent location than earlier traces of theseries, thus precluding contact with andstrengthening of those earlier traces to medi-ate better recall. According to this view, acorrelative behavior that partially indexeswhere S shunts the current input is whetheror not he recognizes the first few groups ofthe string as a recurrence of one experienced

earlier. If S does not recognize the firstfew groups of the altered recurrent string,that means he has shunted it to a new loca-tion and thus has effectively divorced hisimmediate recall of that string from thebenefits of a prior history of experience withit.

An alternative view of matters wouldsuppose that in the context of such serialtasks, S uses only a small number of storageslots or bins that are tied together in serialorder and that all input sequences are re-ferred to these same bins. Suppose in par-ticular that the central processing mechanismswitches the first group (coded unit) of theinput string into the first bin, the secondgroup into the second bin, and so on, allo-cating the sequence of groups or coded unitsin a string to a sequence of storage bins (cf.Conrad, 1965; Neisser, 1967). The suc-cessive bins may be conceived as implicitpositional "stimuli" to which coded responseunits become associated. According to thisview, repetition of an identically structuredstring assures reassignment of the same codedresponse to the same bin, so the strengthen-ing effect on that unit is much like twotrials of A-B, A-B in associative learning.On the other hand, when the group structureis altered between repetitions, entirely dif-ferent coded response units are being as-signed to the successive bins, and the effectat each bin is much like the A-B, A-C para-digm of negative transfer.

Either of these views implies the afore-mentioned results of a repetition effect forRS but not for RC items. Is there a differ-ential prediction of the two views? Thereare doubtless many. One which has ap-pealed to the authors and which is testedbelow is the following implication of the bintheory: If the third (e.g.) group of a stringis repeated exactly, then according to the binhypothesis this constant third group wouldbe associated repeatedly with the third binand thus its recall should improve with prac-tice, irrespective of what is happening in thepreceding and following bins. The con-stituent elements and size (but not the num-ber) of the preceding and following groupsmay vary widely, but the bin hypothesis im-plies that these variations should not affect


the normal improvement in recall of theconstant third group. This implication wastested in Exp. VI.

The reallocation hypothesis leads to amore differentiating set of predictions forsuch "constant chunk" experiments. Recallthat it was assumed that since the string isprocessed sequentially, the central processordecides where to shunt the incoming stringon the basis of recognizing the first informa-tive group or two in the string. To be veryspecific, suppose that if the central processorrecognizes the first group, it shunts the inputstring to that prior storage location whichcaused the recognition; if it fails to recog-nize the first group, then the input string isshunted to a new storage location. Fromthis view of matters, one would predict thata constant first chunk (in an otherwise vari-able string) would improve in its recall overrepetitions since the input string is repeat-edly shunted to the same storage location,there to accumulate trace strength for theconstant portion of the various strings, i.e.,for the constant first chunk. On the otherhand, a constant middle or last chunk wouldnot show improvements in recall because avariable first group would cause each newstring to be shunted to a new location, sothere would be no accumulation of tracestrength for these constant portions at themost recently activated storage location. Incontrast to these differential predictions, thebin hypothesis expects repetition to improverecall of any constant portion of the series,whether it be a first, middle, or last chunk.Experiment VII was undertaken to test thesedifferential predictions.

MethodExperiment VI,—The procedure was similar to

that of Exp. II, except that what recurred wasonly the third group of a five-group string, ratherthan the entire string as in Exp. II. The recurrentthird group was always of Size 4. Other digitsand groupings were arranged around this constantthird group, so that it was preceded and followedby two groups comprised of four total digits, either(2) (2), (3)(1), or (1)(3). These surroundingdigits as well as their groupings were changed be-tween Trials 2, 4, 6, and 8 of each block andonly the third group was recurrent on these trials.On Interspersed Trials 1, 3, 5, 7 of each block, anew noise item was read, usually one that did nothave five groups or a third group of four digits.

The S listened to a tape recording of the phoneti-cally grouped series, then had 10 sec. to write hisrecall in 12 spaces on a recall sheet. He wasurged to report out digits from first to last. Theexperiment contained 10 blocks of eight trials, theblocks separated by 30-sec. rest pauses. A differentfour-digit number was used as the recurrent thirdchunk in each of the 10 blocks. The ^s were 12undergraduate students from the same source asExp. I.

Experiment VII.—This was similar to Exp. VIexcept that the constant chunk was three digitslong and more conditions were compared. Therewere 12 blocks of eight trials, and on Trials 2, 4,6, and 8 of the block some portion of these stringswas recurrent. In different blocks, the recurrentportion in a five-group string was either (a) thefirst group, (6) the fifth group, or (c) the firstand the third group jointly. In the latter case,the second group varied in size (1, 2, or 3) andconstituency over successive presentations of theconstant first and third chunks on the even-num-bered trials of the block. On the odd-numberedtrials of each block, a new noise string was pre-sented and recalled; these usually had more orless than five groups. With 12 experimental blocksand three conditions, each 5 experienced each con-dition in four randomly selected blocks, in ascrambled order counterbalanced over Ss. Also,there was some attempt to counterbalance over SBthe condition to which a particular three-digit num-ber was assigned as the constant chunk. Forexample, "851" was a constant fifth chunk forhalf of the Ss and a constant first chunk for halfof the 5s. This required several tapes to be made.

The S^ were 18 undergraduates taking the sum-mer session course in introductory psychology.They were urged to record the digits in a first-to-last order, although they were not penalized fordoing otherwise.

Results

Experiment VI.—Analyses of mean cor-rect recall over the 10 experimental blocksrevealed no significant performance changeas the experiment progressed, so the 10blocks were pooled for the following analyses.

The bin hypothesis predicts that S's recallof the recurrent third group should improveover its four trials. The focal implicationis with respect to recall of the third groupas a unit, although the exact location of itin 5"s serial output may vary somewhat dueto forgetting or misrecalling the changinggroups that precede and follow the constantthird group. With this in mind, a veryliberal scheme was adopted for scoring re-call of the four ordered digits of the thirdgroup: if the third group as a unit was re-


TABLE 2

PROBABILITY (p) OF CHUNK RECALL ANDMEAN (M) DIGITS RECALLED OF THE

THREE-DIGIT CONSTANT CHUNKS

Condition

First

FirstandMiddle

Last

Statis-tic

PM

PMPM

PM

Trials

1

.471.85

.461.76.11.61

.512.29

2

.541.97

.532.11

.14

.68

.512.32

3

.562.02

.632.30

.15

.72

.442.08

4

.652.22

.612.25

.10

.67

.512.22

LinearF

5.35*3.51

5.10*9.29*nsns

nsns

*p < .05, df = 1, 51.

called with the first digit starting at anylocation from 3 through 7 (true position was5), it was counted as a correct recall of theunit. With this lenient location scoring, theaverage probability of completely recallingthe recurrent unit was .146 on Trial 2, .125on Trial 4, .166 on Trial 6, and .156 onTrial 8. Based on 12 X 10 observationseach, the standard deviations of these pro-portions (assuming independent binomial ob-servations) range around .033, so even themost extreme two proportions do not differsignificantly from each other. With thisscoring in terms of complete recall of theentire unit, there was no evidence of repeti-tion affecting recall of the constant thirdgroup.

A second analysis inquired how manyitems were recalled of the four in the re-current chunk. Using a lenient scoring pro-cedure too complicated to describe briefly,the net result is still the same as above:mean items correctly recalled of the four inthe recurrent chunk were 1.77 on Trial 2,1.67 on Trial 4, 1.72 on Trial 6, and 1.79on Trial 8. Clearly, even with this moresensitive measure, there was no evidence foran improvement in recall over repetitions ofthe constant third chunk. Despite the statis-tical power available in this experiment, itfailed to yield a repetition effect for the con-stant third chunk. This failure disconfirmsthe bin hypothesis since it predicted a strongeffect here, while in fact none was found.

Experiment VII.—The main results ofExp. VII are displayed in Table 2 showingprobability of complete recall of the con-stant chunk and the mean digits recalled inthe constant chunk. There is no apparentimprovement in recall of the constant lastchunk nor in recall of the constant middlechunk. The linear F test for trials in boththese cases is insignificant for both the prob-ability measure and the amount recalledmeasure. On the other hand, there is a smallbut consistent improvement in recall of theconstant first chunk in terms of the prob-ability and amount recalled measures. TheF test for linear trend is significant in threeof the first four rows of Table 2, and in thefourth case the trend is clearly in the rightdirection.

Although recall of the constant first chunkimproved significantly over its repetitions,there was very little change in recall of thevariable nine digits which followed occur-rences of the constant first chunk. Meancorrect recalls of these variable last ninedigits over the four occurrences of the con-stant first chunk were 4.69, 5.11, 4.54, and5.06. These means do not differ reliablyfrom one another.

SummaryIn summary, these two "constant chunk" ex-

periments tend to infirm a bin hypothesis whichsupposes that successive groups of the inputstring are assigned to successive storage binsfrom which the series is read out. That hy-pothesis predicts improvement in recall of afixed group repeatedly reassigned to the samebin independently of whether this group waslocated at the beginning, middle, or end of theseries. The results provide mild support forthe allocation hypothesis since it predictedlearning for a constant first chunk but not fora constant middle or last chunk, and this wasfound. After this experiment was completed,it was learned that Schwartz and Bryden(1966) had performed an experiment with aneven more radical manipulation but tending toa similar conclusion. Using the Hebb-typedesign with digit names read to S's withoutgroupings, these investigators found no im-provement in recall when the last seven of ninedigits were recurrent, but with the first twodigits changed at each recurrence of the lastseven. These and the present results suggest


that the first few digits or first functional groupis acting as an "address" in memory for thedigit string; and when a recurrent portion fol-lows a new address in the input, that string isshunted to a different storage location, so onegets no accumulative benefits in recall for therecurrent portion of the string.

It was said that the allocation hypothesis is"mildly" supported by these data because thereare a few interpretive difficulties. First, since6" was urged to write his recall from first tolast, there is differential output interferenceoperating on the constant first vs. later chunks.Conceivably, S1 might improve his recall of alast constant chunk if he were permitted to writehis recall in any order, i.e., if he wrote firsttiiose portions he might recognize as recurrent.Second, in this interpretation of the allocationhypothesis, the constant middle chunk wasexpected to improve when it was preceded bya constant first chunk. That is, if the constantfirst chunk causes the input string to be re-allocated to the same storage location, then itwas thought that the constant third chunk ofthe string could accumulate trace strength inthat location. But Lines 5 and 6 of Table 2show no improvement in recall of the constantthird chunk in this case. This null learninghere could arise for several reasons. Onepossibility relates it to the variable size (1, 2,or 3) of the second group inserted betweenrepetitions of the constant first and thirdchunks, and this is checked out in Exp. VIII.

The pattern of our constant chunk resultsneeds to be explored with other presentationprocedures, lags between recurrence, recallstrategies, and especially with different typesof materials. Intuitively, it appears unlikelythat serial recall of lists of unrelated words orsentences would be as disrupted as are digitstrings by altering the first few elements ofthe string. If this intuition proves correct, itwould imply that the outcome pattern obtainedhere will be restricted to only those cases inwhich serial lists are constructed by recombin-ing and permuting elements from a very smallvocabulary (e.g., nine digits).

EXPERIMENT VIIIThis final experiment inquires whether the

third chunk will improve in recall when thefirst and third chunk are constant and whenthe second chunk is of constant size. InExp. VII there was no improvement inrecall of the constant third group (cf. Table2) when the second group varied in size

(1, 2, or 3 digits). But this might resultfrom the exact manner in which the inputstring is "overlaid" on the trace of theformer string. If the input string (762)(318) (952) (361) is directly overlaid, ele-ment by element from the beginning, on thetrace of (762) (1) (952) (85) (263), onewould find coincidence of the first threedigits (the first constant chunk), but mis-matching for all the remaining digits. Al-though 952 is the third group in both strings,the variable size of the second group causesmismatching of the 952 elements when thetwo strings are overlaid. 'If this kind ofexact "overlay" process is partly responsiblefor the accumulation of trace strength forparticular digits, then one would expect theconstant third chunk (following a constantfirst chunk) to be learned only when thesecond (variable) chunk is of a fixed size.This single condition was studied in Exp.VIII.

MethodThe procedure was similar to Exp. VI, with

eight blocks of trials in each of which the itemrecurring on Trials 2, 4, 6, and 8 had a constantfirst chunk (Size 3) and a constant third chunk(Size 3). The second group was always of Size 2,but the digits in it varied over the four occurrencesof the other constant portions. The four digitsfollowing the constant third group varied in con-stituency and groupings over the four trials. OnTrials 1, 3, 5, and 7, new noise items occurred,with the group size pattern 323. . . avoided.Grouping was by reading the sequence of numericalnames. Strings were presented by a tape recorder.There were 10 5s from the same source as Exp. I,instructed to write their recall from first to lastdigits.

Results

The results are shown in Table 3 givingprobability of unit recall and mean digitsrecalled (out of three) of the constant firstand constant third chunks. The F test fortrials is significant at p < .01 for bothmeasures for both first and third recurrentchunks in the string.

The learning demonstrated here for theconstant third chunk (following a constantfirst chunk), along with the failure in Exp.VII with a variable-sized second group,lends some credence to the "element-by-element overlay" hypothesis stated above.


TABLE 3PROBABILITY (p) OF UNIT RECALL AND

MEAN (M) DIGITS RECALLED FORTHE FIRST AND THIRD

CONSTANT CHUNKS

Constantchunk

FirstPM

ThirdPM

Trials

1

.501.95

.07

.64

2

.702.37

.201.14

3

.702.50

.251.25

4

.752.57

.281.30

F(3. 27)

6.35*8.65*

7.28*5.39*

*t < .01.

There are several differences between theprocedures in Exp. VII vs. VIII that mightbe relevant; in particular, three differenttypes of recurrent items occurred in Exp.VII, whereas only this one type of recurrentitem occurred in all eight blocks of Exp.VIII. The learning effect on the thirdchunk, indexed by the change from its firstto the last occurrence in each block, wasexamined for the eight successive blocksover the experiment. There appeared to beno particular trend over blocks to theselearning scores. That is, it did not appearthat Ss were coming increasingly to noticeand learn the constant third group as theexperiment progressed. However, such com-parisons may be complicated by various non-specific factors that vary over the course ofthe experiment—practice effects, cumulativeproactive interference, etc. The conservativeconclusion is that the learning of a constantthird chunk, following a constant first chunkand a fixed-sized second chunk, is consistentwith the allocation hypothesis, with the ele-ment-by-element overlay as the mechanismfor accumulation of trace strength.

DISCUSSION

To quickly summarize the results, it hasbeen found that alteration of the group struc-ture of a digit series severely degrades recog-nition of underlying identity, and it preventsthe improvement in recall that normally accom-panies repetition. This was true whether digitgroup coding was directly phonemic or whetherit was produced merely by the location of pausesin the input series. Although recognition of

an altered grouping is impaired, recognition ofan uncoded series was about the same whetherS experienced the same grouping twice or twologically equivalent groupings once each. Whenonly a portion of the digit string constantlyrecurs, its recall improves with repetition ifit is located at the beginning of the string, butnot when it is located at the middle or end ofan otherwise variable string.

These results have been discussed in terms ofthe hypotheses that grouping influences percep-tual coding, that similarity of coding determinesrecognition of identity, and that recognition ofrepetition is normally a concomitant of a prac-tice effect in recall performance. These variousassumptions are summarized in the reallocationhypothesis, which also led to the further experi-ments on uncoded recognition and the locationof a constant chunk in an otherwise variablestring.

One may ask whether alternative hypothesesof serial learning could account for these re-sults. The simple bin hypothesis appears tobe eliminated by the differential results in theconstant chunk experiment. Another alterna-tive is that of simple serial chaining, whichsupposes that serial recall is mediated by for-ward associations between adjacent pairs ofelements. The chaining hypothesis has tworepresentations depending on whether the basicelements associated in pairs are assumed to beindividual digits or phoneme clusters corre-sponding to numerical names of digit groups.The former identification is discredited by thenull learning for RC items in Exp. I and IIsince the same pairwise digit adjacencies re-curred in the RC items. Even if it be grantedthat the pairwise association grows less whenthe pair spans group boundaries, repetition inRC items should nevertheless produce automaticincreases in pairwise associations and serialrecall. But this did not happen. The alterna-tive identification of the basic elements asphoneme clusters converts the chaining hy-pothesis to a more viable theory for these data,despite the fact that it denies the conventionaloperational identification between what 5 islearning and what he is writing down in hisobservable recall.

As stated earlier, the phoneme chaining hy-.pothesis is very similar to the reallocation hy-pothesis and they both imply the RS vs. RCresults of Exp. I, II, IV, and V. However,the results of Exp. Ill, VI, and VII raise afew problems for a simple phoneme chainingtheory. In Exp. Ill, the input phonemes ofRC strings were identical except for the loca-


tion of pauses; the phoneme chaining hypothe-sis can handle the results of Exp. Ill only byassuming implicit receding and rehearsal of thenames of numerical groups suggested by thelocation of pauses. In Exp. VI and VII, andin the more drastic condition studied bySchwartz and Bryden (1966), the constantchunk or chunks should be reflected in a con-stant phoneme portion of a variable stringwithin which the pairwise (phonemic) associa-tions become stronger with repetition. If recallof the constant chunk is in some degree relatedto the strength of its pairwise interassociations(as well as the association leading into theconstant chunk), then recall of it should haveimproved somewhat even when it was locatedin the middle and end portions in the variablestring. But repetition effects on recall werefound only for a constant first chunk.

The authors do not consider these to bemajor objections to the phoneme chaining hy-pothesis, especially since the generality of thepattern of results in the constant chunk experi-ments is not yet established. Thus, the phonemechaining hypothesis should remain as a viablealternative to the metaphorical reallocation hy-pothesis until more discriminating evidence be-comes available.

REFERENCES

ADAMS, J. A. Human memory. New York: Mc-Graw-Hill, 1967.

BERNBACH, H. Stimulus learning and recognitionin paired-associate learning. Journal of Experi-mental Psychology, 1967, 75, 513-519.

CONRAD, R. Order error in immediate recall ofsequences. Journal of Verbal Learning andVerbal Behavior, 1965, 4, 161-169.

FODOR, J., & BEVER, T. The psychological realityof linguistic segments. Journal of Verbal Learn-ing and Verbal Behavior, 1965, 4, 414-420.

GARRETT, M. Syntactic structures and judgmentsof auditory events. Unpublished doctoral disser-tation, University of Illinois, 1965.

HEBB, D. O. Distinctive features of learning inthe higher animal. In J. F. Delafresnoye (Ed.),Brain mechanisms and learning. Oxford: Black-well, 1961.

JOHNSON, N. The psychological reality of phrase-structure rules. Journal of Verbal Learning andVerbal Behavior, 1965, 4, 469-475.

KINTSCH, W. Recognition learning as a functionof the length of the retention interval andchanges in the retention interval. Journal ofMathematical Psychology, 1966, 3, 412-433.

KINTSCH, W., & MORRIS, C. Application of aMarkov model to free recall and recognition.Journal of Experimental Psychology, 1965, 69,200-206.

MARTIN, E. Stimulus recognition in aural paired-associate learning. Journal of Verbal Learningand Verbal Behavior, 1967, 6, 272-276.

McLEAN, R., & GKEGG, L. Effects of inducedchunking on temporal aspects of serial recitation.Journal of Experimental Psychology, 1967, 74,455-459.

MELTON, A. W. Implications of short-term mem-ory for a general theory of memory. Journalof Verbal Learning and Verbal Behavior, 1963,2, 1-21.

MILLER, G., & CHOMSKY, N. Finitary models oflanguage users. In R. Luce, R. Bush, & E.Galanter (Eds.), Handbook of mathematicalpsychology. Vol. 2. New York: Wiley, 1963.

MOSES, L., & OAKFORD, R. Tables of randomPermutations. Stanford: Stanford UniversityPress, 1963.

MUELLER, G. E., & SCHUMANN, F. ExperimentedBeitrage zur Untersuchung des Gedachtnisses.Zeitschrift fur Psychologie und Physiologie derSinnesorgane, 1894, 6, 81-190, 257-339.

NEISSER, U. Cognitive psychology. New York:Appleton-Century-Crofts, 1967.

SCHWARTZ, M., & BRYDEN, M. Retrieval and theeffects of changing elements of a repeating se-quence. Paper presented at the meetings of theCanadian Psychological Association, June 1966.

SHEPARD, R., & TEGHTSOONIAN, M. Retention ofinformation under conditions approaching asteady state. Journal of Experimental Psychol-ogy, 1961, 62, 302-309.

WICKELGREN, W., & NORMAN, D. Strength modelsand serial position in short-term recognition mem-ory. Journal of Mathematical Psychology, 1966,,3, 316-347.

YNGVE, V. A model and a hypothesis for languagestructure. Proceedings of the American Philo-sophical Society, 1960, 104, 444-466.

(Received December 9,1968)

Journal of Experimental Psychology...

Documents

Transcript of Journal of Experimental Psychology...