I

\

'

i

inted

from

Journal of Verbal Leaiininc and Veubai.

Behaviok,

Volum■right

©

1963 by Academic Press Inc.1, Number 4, January, 1963

Printed in li. S. .1.

otNAI

OF

VERBAL

LEARNING

AND VERBAL

BEHAVIOR

1, 269-280 (1962)aI-*

Transfer between Paired-Associate and Serial Learning"

Arthur R. JensenUniversity of California, Berkeley, California

The psychological differences betweenpaired-associate and serial learning are prob-ably far more profound than their formaldifferences would seem to suggest. Indeed,among S-R oriented psychologists the reac-tion to a number of studies which comparethese two forms of rote learning has gen-erally been one of surprise. The principalexperimental investigations of the problemhave been carried out by Young (1959,1961a, 1962). Underwood (1961) recently

reviewed the literature on the problem, in-cluding a number of unpublished studies, andhas indicated their theoretical importance.

the pairs are presented in a different order oneach trial, as is usually done in PA learning.Conversely, a similarly high degree of trans-fer should be expected from a previouslylearned set of PAs, A-B, B-C, C-D, etc.,to learning a serial list, A-B-C-U, etc. Nowthe main point of interest is the fact that theevidence clearly does not support these ex-pectations. Briefly, as of now, the researchpresents the followingpoints.

Transfer from a Serial to a PA List. Onlya negligible amount of transfer has beenfound in going from a serial to a PA list inwhich the S-R elements are common to bothtasks (Young, 1959, 1961a, 1962). For ex-ample, in Young's 1959 study, despite thefact that all the S-R elements were commonto the serial and the PA tasks, there wasonly 8% transfer, a statistically nonsignifi-cant amount. This is typical of Young's laterfindings. A significant amount of transferappeared only in the first few learning trials,but the over-all transfer in these studies waspractically nil. It was also found that 10trials of overlearning of the serial list stilldid not result in significant transfer to the PAlist (Young, 1962).

The traditional S-R conception of seriallearning, which Young (1961a) refers to asthe specificity hypothesis, represents seriallearning as the acquisition of a chain of S-Rassociations. The stimulus for each succes-sive response is assumed to be the item whichimmediately precedes it in the list. Thus,each item in the chain (except the first andlast) is considered to have a double function,serving in turn as a response and as astimulus.

This hypothesis that a serial list is learnedessentially as a chain of paired associates(PA) implies that there should be a highdegree of positive transfer from a previouslylearned serial list, A-B-C-D, etc., to learninga PA list, A-B, B-C, C-D, etc., even when

Transfer from PA to Serial List. Here thepicture is quite different. There has generallybeen found a moderate degree of transferfrom PA to serial learning when the S-Relements are common to both lists. Primoff(1938) found 35% transfer; Young (1959)found 55%. Since even this amount of trans-fer is less than one might expect, consideringthat the PA list was always learned to acriterion of mastery, Underwood (1961, p.

1 This research was aided by a National ScienceFoundation grant to the Center for Human Learn-ing.

2 An abstract of this article was presented at theannual meeting of the Western Psychological As-sociation. San Francisco, California, April, 1962.

269

270 JENSEN

>

a-

.

T

18) has suggested that learning serial listsin the manner of a chain of PAs, though pos-sible, may not be "natural" for the 5. Trans-fer from the PA to the serial list is far lessthan perfect, perhaps, because in goingthrough the serial list as if it were a chainof the previously learned PAs, the 5 mustovercome some tendency to relearn the listin a different manner which is peculiar and"natural" to serial learning. The amount oftransfer may reflect mainly the extent towhich the 5 is successful in maintaining theset for responding to the serial task as if itwere a PA task.

The present experiment is directly con-cerned with this problem. The main questionis whether there is greater transfer from PAto serial learning when it is made relativelyeasy for the S to maintain the set for PAlearning when going to the serial list thanwhen it is made more difficult for the 5 tomaintain his set for PA learning.

One other finding in the literature is quiterelevant to the design of the present ex-periment. This concerns the difference be-tween double-function and single-junction PAlists. In a double-function list each itemserves both as a stimulus and as a response,e.g., A-B, B-C, C-D, etc. In an single-func-tion list the S and R terms are entirelyseparate, e.g., A-B, C-D, E-F, etc. Primoff(1938) was the first to note the great differ-ence in difficulty of learning these two formsof PA task. The 5s required two to threetimes as many trials to learn the double-function lists as they needed for the single-function lists. Young (1961b) performed anexperiment which substantiated Primoff'sfindings and permitted a more generalizedexplanation of the phenomenon in terms ofthe known effects of intralist similarity onrate of learning: the rate of PA learning isinversely related to the degree of stimulus-response similarity. Primoff attributed thegreater difficulty of learning the double-function than the single-function lists to theinhibitory effect of backward associations.

This is also an empirically valid observa-tion. In learning PAs such as A-B, B-C,etc., 5s often give A rather than C as theresponse to B. These first-order backwardassociations rarely occur in serial learning,where the order of presentation is constant,probably because the 5 can remember theitem one position back, since it has so re-cently been given as a response. Conse-quently, when 5 goes from a serial to a PAlist of the double-function type, he is plaguedby a source of errors—first-order backwardassociations—which he did not have to over-come during serial learning, with the resultthat positive transfer appears only on thefirst few trials of serial learning. Soon theerrors of backward association overtake anyinitial transfer effect, and the total amountof transfer at the end of learning is prac-tically zero.

That the inhibitory effect of backwardassociations in the double-function PA list isnot the sole cause of poor transfer fromserial to PA learning, however, is shown byan unpublished study by Young and Benson(cited in Underwood, 1961, p. 19), in whichonly a small amount of transfer occurredeven when the PA task was a single-functionlist.

Findings such as these, which run counterto the expectations of the specificity hypoth-esis, have given rise to two alternative hy-potheses: the compound-stimulus hypothesisand the serial-position hypothesis.

According to the compound-stimulus hy-pothesis, the functional stimulus in seriallearning is not a single item but a compoundof two or more items preceding the responseterm. Thus, the functional stimulus for Din the list A-B-C-D might be BC or ABC.Under investigation this hypothesis has faredno better than the specificity hypothesis.When two items in the learned serial listwere used as the stimulus term in the PAtransfer task, there was still no positivetransfer (Young and Benson, unpublished,cited in Underwood, 1961, p. 19; Young,

271

PAIRED-ASSOCIATE

AND

SERIAL LEARNING

»

i

/

1962). In the latter study there was evennegative transfer!

According to the serial-position hypothesis,originally proposed by Woodworth and Pof-fenberger (1920, pp. 71-72), the functionalstimulus in serial learning is the ordinalposition, or some symbolic equivalent thereof,of each item in the list. The results of ex-perimental investigations of this hypothesisseem to be conflicting [Rehula (UnpublishedPh.D. dissertation cited in Underwood, 1961,p. 21); Jensen and Blank, 1962; Young,1962; Newman and Saltz, 19621. Yet it ap-pears from this evidence that whatever cuefunction serial position per se may possiblyhave, it is probably of minor importance,and Underwood's conclusion that the func-tional stimulus in serial learning has not asyet been identified still seems valid (Under-wood, 1961, p. 22).

The purpose of the present experiment wasto test the hypothesis that the "natural" wayof learning a serial list is psychologicallydifferent from PA learning in that seriallearning does not consist of chaining to-gether successive pairs of S-R elements. Ac-cording to this hypothesis, whatever transferto serial learning results from the prior learn-ing of a derived PA list (aside from general-ized transfer effects such as warm-up, stimu-lus differentiation, and response integration)is due to the s's tendency to carry over the"set" for PA learning to the serial list. Thus,there should be relatively less transfer froma PA list to a serial list when it is madedifficult for 5 to maintain the "set" for PAlearning. The 5 will then tend to learn theserial list in a manner peculiar to serial learn-ing, resulting in less transfer from the priorPA learning than if the PA set were main-tained.

MethodDesign

Two experimental groups were compared forrelative transfer from PA to serial (S) learning.Group PA(Odd) -S first learned a single-functionlist of PAs having the odd-numbered items of the

derived 9-item serial list as the stimulus terms, thus:1-2, 3-4, 5-6, 7-8. Group PA(Even)-S learned firstthe even pairs: 2-3, 4-5, 6-7, 8-9. The second taskfor both groups was, of course, the same serial list:123456789. It was hypothesized that the OddGroup, for whom the serial list begins with a pairalready learned (1-2), should find it easier to main-tain the PA set than would the Even Group. Thus,if a different strategy exists for serial than for PAlearning, the Even Group should be more pronethan the Odd Group to lose the set for PA learningand to adopt the strategy of serial learning.

A Control Group learned the serial list first. Inorder to assess transfer from the serial to the PAlist, some of these 5s then learned the odd PA list.

Transfer due to response integration and acquisi-tion per se was minimized by composing the serialand PA lists of colored forms, which in previousexperimentshave been shown to attain high responseavailability as soon as they have been described inthe preliminary instructions to the 5. Thus, prac-tically all the learning that occurs with these stimuliinvolves only the associative phase of PA or seriallearning. Since a preliminary study indicated thatthe PA task would have been too difficult for most5s if the pacing interval had been the same as in theserial learning (3 sec), and since it was desired tohave 5s overlcarn the first task to a rigorous cri-terion and yet not be unduly fatigued or suffer amotivational slump by the beginning of the secondtask, it was decided to use sell-pacing in the PAtask.

An aspect of procedure that is probably uniquein PA-serial transfer experiments is that the 5shere were clearly informed of the method by whichthe PA list was made up from thc serial list (andvice versa), and in going from the first to thesecond task they were explicitly instructed to try touse the S-R connections they had acquired in thefirst task. The experiment was represented to the 5sas a test to determine how well they could transferwhat they learned in one situation to another. Thisprocedure more or less insured that any failure oftransfer to occur could not be attributed to thes's failure to perceive the possibility for transferfrom the first to the second task.

SubjectsThe 5s were 171 juniors and seniors (35 men and

136 women) recruited from an introductory coursein educational psychology at the University ofCalifornia. Three women were eliminated from theexperiment for refusing to persist in the first taskuntil they attained the criterion. Of the remaining168 ss, all attained the criterion on both the firstand second tasks. The number of 5s in each treat-

272 JENSEN

«

i

-

ment group was as follows: 25 in Group S—PA(Odd); 54 in Group PA(Odd)—S; 54 in GroupPA(Even)— S. The serial learning data from acontrol group (N = 35) which learned only theserial list was combined with the serial data ofGroups S—PA(Odd), and these combined data wereused as the control (with N — 60) against whichrelative transfer was measured in the two groups(Odd and Even) going from PA to serial learning.

ProcedureThe PA and serial lists were presented by two

different apparatuses. The 5 sat a few feet in frontof the apparatuses and could change from the ap-paratus used in the first task to that used in thesecond merely by turning 90 degrees in a swivelchair.

The stimulus items common to all tasks werecolored forms—squares, triangles, and circles coloredred, yellow, and blue. In the serial list from whichthe PAs were composed, each form appeared oncein each of the three colors, and items of the sameshape or the same color were never adjacent in thelist. The color-forms appeared approximately 2 in.in size on the screen and the colors were vivid.

Instructions. All 5s were told five things: (a) allthe items to be learned would consist of color-forms(as described above) and items of the same shapeor the same color would never be adjacent in thelist; (6) the 5 was to learn by the anticipationmethod, by saying, for example, "red triangle," "bluesquare," etc., and guessing was encouraged from thefirst trial on; (c) S had to learn to a criterion ofthree consecutive errorless trials; (d) the serial taskwould be paced at a 3-sec. rate and the PA taskwould be unpaced (or self-paced)

;

(e) the secondtask would be made up of the same S-R connec-tions learned in the first task, for this was a testto see how well the 5 could transfer the first-tasklearning to the second task.

The procedure for PA learning or serial learningwas exactly the same whether it was the first taskor the second.

PA Task. The paired stimuli, first one and thenthc two together, appeared in two side-by-sideground-glass windows about 2 in. apart. On eachtrial the stimulus item was first presented alone inthe left-hand window until the 5 made his anticipa-tion and then the response term appeared in theright-hand window; the two items were displayedsimultaneously for approximately 2 sec. and thenthe screen went blank for approximately 2 sec. be-fore the next stimulus item appeared. The intervalbetween the S and R terms was governed by eachs's own rate of responding. The order of presenta-tion of the four PAs was different for every 5 and

was random on every trial, with two exceptions:(a) the same pair was never repeated in immediatesuccession; and (6) every pair was presented once

within each set of four presentations. The Ss learnedto a criterion of three successive errorless trials, eachtrial consisting of all four PAs.

When the PA task came

first,

immediately onattaining the criterion the S was told to turn tothe other display panel and was reminded to try totransfer the PAs he had just acquired to the seriallist about to be presented. The 5 began anticipatingthe serial list on its first presentation.

Serial Task. The 9-item serial list was alwayspreceded by a green light as the signal for the firstanticipation. The items were presented at a 3-sec.rate, with a 6-sec. intertrial interval. The 5s learnedto a criterion of three successive errorless trials.

When the serial task came

first,

immediately uponattaining the criterion, the 5 was told to turn to theother display panel and to try to transfer the S-Rconnections he had just acquired to the PA list.

ResultsSince the primary concern was the relative

amounts of transfer in the two experimentalconditions [PA(Odd)—S and PA(Even)—S], no attempt was made to equate the ex-perimental and control groups for generalizedtransfer effects such as warm-up and learning-to-learn, as would be necessary in order toestablish the absolute amount of specifictransfer. Since the two experimental groupsdid not differ in the first-task learning (PA),it can be safely assumed the generalized prac-tice effects were the same for both groups.

Two measures of performance were used:number of trials to criterion (of three suc-cessive errorless trials) and the percentage oferrors (100 X the ratio of errors to all op-portunities for error).

The principal results of the experiment aresummarized in Tables 1 and 2.

Transfer of Serial-Position Effect to Paired-Associate LearningAs shown in the lower half of Table 1,

Group S-PA(Odd) learned faster than thecorresponding control group (PA Odd) ; thedifferences both for trials to criterion and forpercent errors are significant beyond the .001level (t = 3.34 and 5.74, respectively).

273

PAIRED-ASSOCIATE

AND SERIAL LEARNING

i

a

i

Trial 1 errors Trials to criterion Per cent errors

SDGroup

1/

SD M SD M

Serial Learning7.73 43.49 7.31Control 7.83 1.10 24.98

PA(Odd)—SPA(Even)—S

5.70 1.49 16.50 6.91 36.69 9.9042.87 9.017.02 1.48 22.48 8.40

PA Learning

17.48 5.97 43.99 12.80PA(Even)PA(Odd)S—PA(Odd)

12.2118.54 7.28 48.0212.24 7.83 30.97 12.02

The relative amounts of transfer to PAlearning for the individual pairs is a func-tion of the serial position held by the itemsduring serial learning, as shown in Fig. 1.Note how perfectly the serial-position curveemerges in the transfer to the PA learning,even though the pairs were always presentedin a different order. The control group pairswould form a practically straight horizontalline in Fig. 1. Analysis of variance yielded aGroups X Positions interaction significantbeyond the .001 level. Apparently by the endof serial learning the items differ in associa-tive strength according to serial-position, withthe earliest learned items being the most over-

S-R

Pair in

Serial

List

Fig. 1. Mean percentage of errors made on eachof the pairs in the PA transfer task. The numberson the abscissa indicate the serial positions occupiedby the S and R members of thepair during the seriallearning which preceded the PA task, in which theS-R pairs were presented in a random order.

learned and consequently having the greatestassociative strength.

It seems clear that ss, after learning (oroverlearning) a serial list, are to some extentable to respond to the appropriately derivedpaired-associates as if the serial list had beenlearned in accordance with the specificity hy-pothesis. The basic question, however, iswhether the acquisition of a chain of S-Rassociations is essential to serial learning oris merely incidental learning.

Transfer from Paired-Associate to SerialLearningPA Learning. As shown in Table 1, the

means and standard deviations of the twotransfer groups, PA(Odd)—S and PA(Even)—S, do not differ significantly on either ofthe measures in the first-task PA learning.The Odd and Even pairs are quite equiva-lent. For trials and for percent errors the t'sare 1.00 and 1.66, respectively.

Serial Learning. Group PA(Odd)—Sshowed significantly faster learning of theserial list than Group PA(Even)—S. Fortrials to criterion, t — 4.01, for percent errorst = 3.36; for both, P < .001. The OddGroup differs significantly from the ControlGroup at the .001 level (t = 6.11 and 4.10for trials and percent errors, respectively),while the Even Group does not differ signifi-cantly from the Control Group (t = 1.63and < 1.00 for trials and percent errors, re-spectively).

Table 1Summary

of

Data on Learning under the Different Conditions

274 JENSEN

4

l

Table 2Results for Items PreviouslyLearned and Items Not Previously Learned

in Transfer from PA to Serial List

As shown in Table 2, transfer is con-siderably greater for the associations specifi-cally learned in the PA task than for thosethat were not a part of the PA task, andthis is true for both the Odd and EvenGroups. The Odd Group showed significantlyfaster learning than the Even Group bothfor items previously learned and for itemsnot previously learned (in terms of numberof errors to criterion, t = 5.62, P < .001,for previously learned items, and t = 3.98,P < .001, for not previously learned items).

First-Trial Data. Is the transfer from PAto serial learning an immediate effect or is itmore a "savings" effect which shows up onlylater in the course of serial learning? Com-parison of the first-trial data of the transferand control groups, shown in Table 1, pro-vides the answer. (The first-trial responsesof the Control Group were pure guessing,since 5s were required to begin anticipatingon the very first presentation of the list.) Theamount of transfer on the first trial is com-parable to that for the entire course of learn-ing. Again, the Odd Group shows significantlymore transfer than the Even Group (t =4.57, P < .001). Note, however, that on thefirst trial the Even Group is also significantlybetter than the Control Group, but this ad-vantage is mostly lost in later trials. As one

would expect, the only significant transferon the first trial occurs on the associationspreviously learned in the PA task.

Failures to Respond. Did the Odd Grouptend to carry over the PA "strategy" intoserial learning to a greater extent than didthe Even Group? That is, did the Odd Grouptend more than the Even Group to regardthe serial list as a chain of S-R pairs? If so,it would seem reasonable to expect a rela-tively greater percentage of failures to re-spond for the Odd Group on those itemswhich were stimulus terms in the PA task.The items that preceded each of these in theserial list, having been response terms in theTA list, had thus never functioned as stimuli.Some change in the s's "set" would seemnecessary for these terms to elicit anticipa-tions, as required by the serial task. Thequestion, then, is, does the Odd Group change"set" less readily than the Even Group?

Since the Odd and Even Groups differ intotal error rate, it is necessary, to examinethis point properly, first to determine thepercentage of all errors that are failures torespond. This was done separately for eachposition in the serial list, except for theitems that never appeared in the PA list, i.e.,the last item of the serial list for the OddGroup and the first item for the Even Group

Trial 1 errors Errors Per cent errors

1/ SDGroup SD SDPreviously Learned Items

Control (Odd)Control (Even)

3.413.56

0.71 41.120.64 42.32

15.5316.85

46.8047.90

11.438.61

PA(Odd)—SPA(Even)—S

1.502.33

1.17 17.021.17 34.72

12.0719.50

29.3043.63

13.5412.67

Not Previously Learned ItemsControl (Odd)Control (Even)

4.424.27

0.74 45.48 17.500.93 43.83 16.26

41.4539.97

7.428.04

PA(Odd)—SPA(Even)—S

4.204.69

0.83 28.22 16.310.57 41.57 18.14

42.1444.12

11.196.02

275PAIRED-ASSOCIATE AND SERIAL

LEARNING

*

*

a

a-

(these items were omitted from this analysis).From these data, on which the Odd and EvenGroups were now, in effect, equated for totalerrors at each serial position, it was possibleto compare the percentage of failures on theitems that had been stimulus terms and onthose that had been response terms in thePA list. The PA stimulus terms (Positions1, 3, 5, 7) for the Odd Group account for60% of the failures to respond in seriallearning; the response terms account for40%- The PA stimulus terms (Positions 2,4, 6, 8) for the Even Group account for 48per cent of the failures in serial learning;the response terms account for 52 per cent.(When the percentage of failures is not cor-

rected for total error rate, the PA stimulusterms for the Odd and Even Groups accountfor 68% and 53 % of the total failures, re-spectively.)

These results are highly consistent withthe notion that the Odd Group tends to re-gard the serial list as a chain of PAs andtherefore experiences relatively greater dif-ficulty in making responses to the responseterms carried over from the PA task. The5s tended simply to wait for the next stimu-lus item, as they had done in the PA learning,rather than to anticipate the next item, asrequired by the serial task. The Even Group,on the other hand, showed practically nodifference in failure rate between the PAstimulus and response terms in serial learn-ing.

Serial-Position Effects. The serial-positioncurves show the differences between the Oddand Even Groups most strikingly. The curvesin Fig. 2 plotted in terms of mean errorsat each position, clearly show the effects ofthe prior PA learning for the Odd Group,while the Even Group produced an almosttypical serial-position curve. (Since the curvefor the Control Group was quite typical,it was not entered in this graph.) It appearsthat the Even Group tended to lose its pairedassociations in going to the serial list, whilethe Odd Group was able to some extent to

Serial

Position

Fig. 2. Serial-position curves showing mean errorsat each position for the serial transfer groups PA(Odd)—S and PA(Even)—S. Thc S-R connectionsthat had been learned in the prior PA task areindicated by the letter L, and by the solid (Odd)and dotted (Even) lines connecting the numbers onthe abscissa.

transfer the PA connections to the seriallearning. In order to get a closer look at thisphenomenon, it was decided to plot the serial-position curves in terms of each one-fourthof the trials to criterion. Figures 3, 4, and 5show the serial-position curves of the Control,Odd, and Even Groups, respectively, whenthe curves are plotted in terms of the meanerrors made in each quarter.

It is evident in Fig. 4 that the serial-posi-tion curves of the Odd Group continue toreflect the marked transfer effects of thepreviously learned items through at leastthe first half of the. trials-to-criterion. In theEven Group (Fig. 5), on the other hand, theeffects of the PA learning are reflected onlyin the first fourth of the trials, after whichthe serial-position curve begins to assumea more or less typical appearance. By thelast fourth of learning, the curves of bothgroups are fairly typical serial-position curvesand resemble closely that of the ControlGroup (Fig. 3). These results seem to sug-gest that the Odd Group may show greatertransfer mainly because they were better

14 | 1

Even P-A's (N = 54)~\12 - \

£ 10 - /

\

I"" / Ax/A \A/ Vy Odd P-As (N= 54)

L= Learned

Connection

oLIII I I I II 2 ----3 4-- --5 6-—7 8 — -9

276 JENSEN

4-

i

4

.

Fig. 3. The serial-position curves, in terms of mean errors, for the Control Group, plotted for each onefourth of the trials to the criterion of three successive errorless trials.

Fig. 4. The serial-position curves, in terms of mean errors, for the PA(Odd)—S group, plotted for eachone-fourth of the trials to thc criterion of three successive errorless trials. The lines connecting the numberson the abscissa indicate the S-R connections learned in the PA task.

Fig. 5. Thc serial-position curves, in terms of mean errors, for the PA(Even)—S group, plotted foreach one-fourth of the trials to the criterion of three successive errorless trials. The lines connecting thenumbers on the abscissa indicate the S-R connections learned in the PA task.

able to retain the PA set in the serial learn-ing. The fact that they were not completelysuccessful in this is shown by the less thanperfect transfer even on the previouslylearned items and by the emergence of rathertypical serial-position curves in the last halfof learning.

The "Fate" of a Single, Learned PairedAssociate. Was the Even Group more or lessforced into abandoning the PA associationsin order to be able to learn the list in ac-cordance with some different strategy peculiarand "natural" to serial learning? To get atthis, learning curves were plotted for singleitems. Figure 6 shows a typical set of suchcurves. It shows the percentage of errors (orthe percentage of 5s who fail to give thecorrect response) made on each of the first 11trials (which is approximatelyhalf the meannumber of trials to criterion) on the itemnearest the middle position in the list whichhad been learned as a response in the PAtask. Thus, for the Odd Group is shown thelearning curve for Position 6 (which hadbeen learned as the response term of the

12 3456 789 10 IITrials

Fig. 6. Learning curves, in terms of percentageof errors (or percentage of 5s making errors) oneach of the first 11 trials of serial learning, for theitem nearest the middle of the list which had beenacquired in the PA task. The Control Group curverepresents the mean percentage of errors on eachtrial for Positions 5 and 6 in the serial learning. Thecurve for the PA(Odd) —S group is for Position 6,which had been learned in the PA task as the re-sponse to the item in Position 5. The curve for thePA(Even)—S group is for Position 5, which hadbeen learned in the PA task as the response to theitem in Position 4.

277

PAIRED-ASSOCIATE

AND

SERIAL LEARNING

r

<

a

pair 5-6 in the PA task). The curve of theEven Group is for Position 5 (the responseterm of the paired-associate 4-5). The curveof the Control Group is the mean of Posi-tions 5 and 6. The results for the ControlGroup, of course, reveal a typical learningcurve, with a fairly regular decrease in er-rors as a function of trials. The Odd Groupshows an advantage from the very first trial,but there is practically no improvement inperformance on this item through approxi-mately the first half of learning. (This isreflected, too, in Fig. 4.) The Even Groupshows an initial advantage, but it quicklyfades, and by Trial 6 they perform no bet-ter than the Control Group.

These results are clearly in accord withthe hypothesis that the PA strategy breaksdown for the Even Group, which then aban-dons the previously learned PA connectionsand learns the list in a fashion peculiar to

serial learning.

Correlations Between Serial and Paired-As-sociate LearningIf serial and paired-associate learning in-

volve essentially the same kind of learningprocess, one should expect a fairly high cor-relation between individual differences inthe two forms of learning, especially whenthe learning materials are the same in bothmodes of presentation. Table 3 presents the

Table 3Correlations (Pearson r) between PairedAssociate and Serial Learning, Based on

Total Errors to Criterion"

« Note: for a (2i df) a correlation of .396 wouldbe significant at the .05 level; for b through g(52 df) the required value is .268.

correlations between PA and serial learningfor the various groups and conditions in thisstudy. None of the correlations is significantlygreater than zero at the .05 level, although allare positive, which suggests that there mightbe some true correlation, albeit slight. Thevariance the two tasks share in common, how-ever, seems surprisingly small, consideringtheir formal similarity. Scores based on thepreviously learned items show no higher cor-relations than items not previously learned.

It is unlikely that these low correlationsare due to low reliability of the learningmeasures. Some idea of the reliability ofthe serial learning measures is suggested bythe correlations between error scores onthe odd and even items in the serial list.These "split-half" reliabilities for the Odd,Even, and Control Groups were .88, .89, and.92, respectively. The "split-half" reliabili-ties of the paired-associate learning was ob-tained for the Odd Group and Control Groupby correlating the errors made on pairs 1-2and 3-4 with errors on 5-6 and 7-8; for theEven Group the errors on pairs 2-3 and 4-5were correlated with pairs 6-7 and 8-9. Thereliabilities thus obtained for paired-associatelearning in the Odd, Even, and ControlGroups were .90, .84, and .89, respectively.

Discussion

These findings, along with those reviewedfrom the literature, invite the following gen-eralizations and speculations, (a) Serial andpaired-associate learning depend upon differ-ent processes or strategies of learning whichpsychologically have little in common, ifpaired-associate learning is viewed as theformation of S-R associations, (b) A seriallist is not normally learned as a chain of S-Rassociations, (c) The slight transfer that oc-curs from serial to PA learning, when theyformally have S-R elements in common, maybe due to incidental learning of the PA con-nections in the serial list rather than to afundamental similarity between serial andPA learning. In the present experiment there

Tasks

(a) S—-PA(Odd) .295(b) PA(Odd)—S (Total List)(c) PA(Odd)—S (Learned Items)(d) PA(Odd)— S (Not-learned Items)

.035.035.034

(p) PA(Even)—S (Total List)(/) PA(Even)—S (Learned Items)(g) PA(Even)—S (Xot-learned Items)

.207

.129

.265

278 JENSEN

■a

>

a

Va

is no way of evaluating the degree of trans-fer attributable to generalizedtransfer effects,but even if there were a good deal of suchtransfer, the remarkable thing, consideringthe conditions of the experiment, is the rela-tively small degree (about 50%) of over-alltransfer from the serial to the PA list. Theconditions for producing specific transferseemed close to maximal: 5s were told theywere supposed to transfer; only four single-function PAs were derived from the seriallist; the serial list was overlearned to a cri-terion of three successive correct trials; andthe PA presentation was unpaced, so that the5 could have run through the serial list in hismemory to find the terms that followed eachstimulus item in the PA list. Under theseconditions, if what is learned in serial andPA learning is essentially the same, it seemssurprising that transfer was not close to 100per cent, (d) In transfer from PA to seriallearning, 5s who are enabled to carry overtheir set for PA responding into the serialtask learn the serial list faster than do 5sfor whom the congruence, or lack of it, be-tween the PA and serial list is not such as tofacilitate continuance of the PA response set.

Other evidence that different processes areinvolved in PA and serial learning are thenegligible correlations between individual dif-ferences in the two forms of learning andthe difference in difficulty of learning thePA and serial lists. The latter point can onlybe surmised from the present data, sincethe number of items in the PA and seriallists differed as well as the pacing of theirpresentation. But consider the following facts,based on first-task (control group) data:On a 9-item serial list presented at a 3-sec.rate, 5s required on the average approxi-mately 25 trials to attain criterion, whileon a 4-pair paired-associate list (which in-volved learning only four-ninths as manyconnections as in the serial list) presentedat a self-paced rate, 5s required approxi-mately 19 trials to attain the same criterion.[In a previous experiment (Jensen, 1962b)

in which 5s learned the same kind of 9-itemcolor-form test with an unpaced rate of pres-entation, the average number of trials neededto attain one errorless trial was between 3and 4. | Since the PA task was the single-function type, there was not the added inter-ference which would have resulted had adouble-function PA list been composed fromthese color-form materials. It is doubtful ifthe majority of 5s could have mastered adouble-function list in any reasonable amountof time. If the same kind of learning weregoing on in the two forms, why should theydiffer so greatly in difficulty?

But the really central question is exactlyhow serial learning differs from PA learning.To use Underwood's term (1961, p. 18), whatis the "natural" manner in which a seriallist is learned? One possible answer is sug-gested in the writer's tentatively formulatedtheory of serial learning (Jensen, 1962a).The hypothesis is not essentiallyan S-R con-ception of serial learning; it even suggeststhat a search for the "functional stimulus"in serial learning may be a vain pursuit. Thetheory holds that the "natural" process oflearning a serial list consists of "attaching"responses to an anchor point (the first item,the intertrial blank space, or the signal pre-ceding the first list-item) in both a forwardand backward direction. What the 5 is doingis not linking up a chain of S-R associations,with each item acting as the stimulus forthe next, but is integrating a number of re-sponses. All that is meant by the term inte-gratedresponses is that the response elements(e.g., words or nonsense syllables) can beemitted by the 5 in a particular sequencewithout their being individually dependentupon specific eliciting stimuli or cues. Thisapplies to either external stimuli or to re-sponse-produced stimuli. An example in themotor realm is a pianist's execution of arapid passage of notes. There is a definitesequence of finger movements, but it. is knownthat these movements can take place sorapidly as to make it impossible that they

279

PAIRED-ASSOCIATE

AND SERIAL

LEARNING

►

>

*

could be guided by external stimuli such asthe printed score or the preceding sounds, orby proprioceptive feedback from the fingermuscles. This complex response apparentlyissues from some centrally integrated process,and its sequential elements are not dependentupon eliciting stimuli for their execution.Another example of centrally integrated re-sponse elements is the immediate memoryspan: an 5 can make a series of responses,such as repeating a sequence of digits, aftera single presentation of the series. Thereseems to be no specific stimulus for eachdigit. The units of a serial list can be thoughtof as being integrated in the same sense thata shorter series, comprehended by the memoryspan, is integrated. Through repeated trialsthe S can integrate a longer series of responseelements than can be comprehended by hismemory span. The items of a serial list mightbe conceived of as serving not essentially asthe stimuli for anticipations, but as rein-

forcers of the s's responses. Of course, theitems also provide the s's repertoire of re-sponses that become integrated as the seriallist. Thus, psychologically, the list is not

composed of S-R connections; the items neverreally take on functional stimulus propertiesin the sense that they must do in PA learn-ing. In PA learning the largest unit of inte-grated response is only two items, and thuseach S-term is crucial to performance. Inserial learning, on the other hand, the 5,after a certain number of trials, is able onrequest to recite the serial list without beinggiven any formal stimulus whatsoever. Thelist has become an integrated response with-out the need for specific stimulus cues alongthe way. An heuristic analogy might be to

consider the groove in a phonograph record,which for its initial formation depends upona sequential "stimulus" input, but which onlater playings, once the stylus has been setin the groove, gives off the sequence of tones

without any sequence of "stimuli" being in-volved at all. Certainly each successive toneis not the "stimulus" for the next. A similar

sort of thing may be true for human sequen-tial acts, such as playing a piece on the piano,assembling an apparatus, or learning a list ofnonsense syllables. A pattern for an integratedresponse is laid down in the nervous system,likened to the groove on the phonograph rec-ord, rather than a sequence of discrete S-Rassociations, which might be likened to thechain reaction in a row of dominoes when thefirst one is tipped over. This conception ofserial learning should suggest many experi-ments capable of further testing its validity.

Summary

It was hypothesized that serial learningtakes place by a different process or strategythan PA learning and does not normally con-sist of the chaining together of S-R connec-tions. It was also hypothesized that specifictransfer of S-R pairs from PA to serial learn-ing would be facilitated under conditions thatmake it relatively easy for the 5 to carryover his set for PA learning into the serialtask as compared with conditions that makeit relatively difficult to maintain the PA set.To test this hypothesis two experimentalgroups learned a PA list followed by a seriallist with common S-R elements. A controlgroup learned only the serial list. The twoexperimental groups, labeled Odd and Even,learned different pairs of items derived fromthe serial list. The Odd Group learned pairs1-2, 3-4, 5-6, 7-8; the Even Group learnedpairs 2-3, 4-5, 6-7, 8-9. The second task con-sisted of learning the serial list 12 3 4 5 6 78 9. The Odd Group, for whom the set forpaired-associate responding in serial learningwas facilitated by the fact that the first itemsin the serial list had already been learned asa paired associate, learned the serial list sig-nificantly faster than the Control Group.Though the Even Group showed some ad-vantage over the Control Group in the firstfew trials of serial learning, this advantagedisappeared completely by the sixth orseventh trial. This was true even of theadjacent serial items that had been pre-

280 JENSEN

;

viously learned as paired associates. This re-sult was interpreted as being due to a lossof the PA set and the adoption of a strategypeculiar to serial learning. Other evidencethat quite different processes are involved inserial and PA learning was the lack of asignificant correlation between individual dif-ferences in the two forms of learning.

It appears that a serial list is not learnedas a chain of S-R connections and that whatthe 5 learns in the serial task is somehowquite different from what he learns in the PAtask, even though both tasks formally havethe same associative elements in common.An hypothesis was suggested to account forthese findings, viz., that in PA learning theS-terms serve primarily a cue or stimulusfunction, while in serial learning the subjectintegrates a number of responses (suppliedby the items in the list) and the items serveprimarily as reinforcers without acquiring acue function, except possibly by incidentallearning when the list is overlearned.

References

Jensen, A. R. An empirical theory of the serial-position effect. J. Psychol, 1962, 53, 127-142. (a)

OJensen, A. R. Temporal and spatial serial-posit

effects. Amer. J. Psychol., 1962, 75, 390-400.Jensen, A. R., and Blank, S. S. Associations v ®

ordinal position in serial rote-learning. Canau.J. Psychol, 1962, 16,60-63.

Newman, S. E., and

Saltz,

E. Serial position as acue in learning. Amer. J. Psychol., 1962, 75,102-108.

Primoff,

E. Backward and forward association asan organizing act in serial and in paired as-sociate learning. /. Psychol., 1938, 5, 375-395.

Underwood, B. J. Stimulus selection in verbal learn-ing. Paper for the Second Office of Naval Re-search Conference on Verbal Learning. June,1961.

Woodworth, R.

S.,

and Poffenberger, A. T. Text-book of experimental psychology. Mimeographeded. New York: Columbia Univer., 1920.

Young, R. K. A comparison of two methods oflearning serial associations. Amer. J. Psychol.,1959, 72, 554-559.

Young, R. K. The stimulus in serial verbal learn-ing. Amer. J. Psychol., 1961, 74, 517-528. (a)

Young, R. K. Paired-associate learning when thesame items occur as stimuli and responses. J.exp. Psychol, 1961, 61, 315-318. (b)

Young, R. K. Tests of three hypotheses about theeffective stimulus in serial learning. /. exp.Psychol, 1962, 63, 307-313.

(Received August 8, 1962)