BOWER #3

43 d»rt*rj Unfvwsfty LiopaneeDspt. of SpwMi C<*fec*on6

CofiBoxFof.

H-

Fol. Title

journal

of

verbal learning and verbal behavior 8, 323-343 (1969)

4

4

i

%.

Hierarchical Retrieval Schemes in Recallof Categorized Word Lists 1

Gordon H. Bower, Michal C. Clark, Alan M. Lesgold, and David Winzenz

Stanford University,

Stanford,

California 94305

These experiments investigate the effects of hierarchic organization of word-lists upontheir free recall. 5s recalled nested category lists presented either randomly or in a hier-archically organized manner. Recall was 2-3 times better with the organized presentation.Later experiments showed this effect (a) was similarwith associative as well as conceptualhierarchies, (b) was attenuated with recognition tests of memory, and (c) could not beaccounted for by associative "guessing." Another experiment demonstrated retroactivefacilitation in recall of List 1 when List 2 contained the hierarchic superordinates of thewords on List I. Analyses suggest that the hierarchic principle was used as a retrieval planfor cuing recall, with generated candidates monitored fortheir list membership before beingovertly recalled.

The followingexperiments wereundertakento demonstrate the influence of structuralorganizationupon the free recall of conceptualword hierarchies. The studies were initiatedin the belief that previous free recall experi-ments have investigated only relatively weakmanipulations of this structural variable, andthe further belief that it should be possible toarrange for a much more potent demonstra-tion.

order, whereas in random presentation theseveral instances of each category are sepa-rated by many intervening items in theexposure list. Psychologically, blocking ofinstances at input is more likely to lead S todiscover the superordinate category(cf. Woodand Underwood, 1967) and thus it shouldmediate better recall. Although blocked listsare often recalled better than random lists,the effect has often been surprisingly small.For example,in theCofer et al. study, blockingproduced an average recall advantage of only13% overarandomly ordered list. In Cohen'smore recent and extensive studies, with longlists of many categories, there has been noeffect at all of blocking vs.random input upontotal recall.

This research is most relevant to previousstudies of freerecall of categorized word lists.If the list consists of several instances ofseveral taxonomic categories (e.g., animals,colors, occupations, etc.), it will be betterrecalled than a comparable list of unrelatedwords; moreover, in his recall, S tends togroup together words from the same category.A number ofvariables inherentin this situationhave been studied (e.g., Cofer, Bruce, andReicher, 1966; Cohen, 1966). One of imme-diate relevance to our studies is the presen-tation-order of a categorized word list. Inblocked presentation, the instances of a givencategory are presented in adjacent temporal

If one believes that structural informationabout the input list is an important componentof S's ability to recall it, then the small ornon-existent effect of blocking in these priorstudies is rather disheartening. However,the possibility remains that the structural-organization variable hasbeen only weaklyandineffectively manipulated in the prior studies.In consequence, one searches for a morepotent way to manipulate this variable, andthis is what we have done in the followingexperiments.

1 This research wassupported by a grant (M- 13950)to the first author from the National Institute ofMental Health.

11 323

© 1969 by Academic Press Inc.

324

BOWER

ET AL.

>

*■'

i

LEVEL

Fig. 1. The "minerals"conceptual hierarchy.

We have made three obvious changes fromthe prior studies of category blocking. First,the list to be recalled contains not onlyinstances of a category, but also the categorylabel itself. Ifrecall ofthecategorylabel servesas a cue for recall of its instances (cf. Tulvingand Pearlstone, 1966), then the presence ofthe category name on the recall list shouldboost the recall of the structured material.

Second,

the word list is selected to be ahierarchically organized set of nested cate-gories, so that the "instances" ofa high-levelcategory serve in turn as superordinates forstill lower-level instances. We have constructedabout eight of these conceptual hierarchies,one of which(for minerals) is shown in Fig. 1.The four different "levels" of the hierarchicaltree are indexed by the left-hand column ofnumbers.

Such word trees have an obvious structuredefined generally by the "class inclusion"relationship, and there would probably befairly general agreement on the appropriate-ness of the words at different levels of the tree.Surprisingly, however, we have found thatnaive Ss have considerable difficulty in tryingto generate such trees given only the level-1word and a general characterization of thetarget hierarchies to be generated (cf. ourExp. V later). The notion of "levels of classinclusion" turns out to be a terribly vague andimprecise concept when applied to natural

language, and in actuality there are a verylarge number ofplausible options orbranchingpoints in generating such a hierarchical tree.Because of the open, nonexhaustive characterof such conceptual word hierarchies, the oneactually used can be considered as only oneof many plausible trees that could have beendeveloped from the level-1 word. In formalterms, the trees used in these experiments varyconsiderablyfrom one another in the numberof nodes at each level, and the use of nounsversus class-restricting adjectives at the higherlevels. Many of the words have ambigu-ous meanings, so that the intended senseof the word would be established onlyby seeing it in relation to its hierarchicalcontext.

The third change we have made from theprior studies is in the method of presenta-tion. The prior studies have typicallypresented the lists one word at a time, whereaswe present a complete set of words all at oncefor prolonged exposure. The one-at-a-timemethod has the advantage of operationallyequating exposuretime to each word, whereasthe complete-presentation method can onlyequate total time for the entire set of words.However, themethod ofcompletepresentationmakes it easier for S to discover more struc-tural information about the input list thandoes the one-at-a-time method, and theformerhas been used for this reason.

325HIERARCHICAL RETRIEVAL SCHEMES

i

to

Experiment IOur first experiment is in essentials a simple

comparison of free recall of hierarchical wordlists thatarepresented in ablocked as opposedto a randomized fashion. Four word hier-archies were learned concurrently. The S's inthe Blocked condition were exposed to thefour conceptual hierarchies organized invertical trees as shown in Fig. 1. For S's in theRandom condition, the same words werethoroughly scrambled, then assigned ran-domly to the nodes of four spatial trees. Thespatial tree seen by Random Ss had noapparent conceptual significance since thewords located above and below a given node-word bore no obvious conceptual relation tothe node.

A subsidiary factor that was varied ortho-gonal to the Blocked versus Random factorwas Progressive parts versus Whole presen-tation of the hierarchies. In the Progressiveparts condition, the level-1 and -2 words werepresented on Trial 1, levels 1, 2 and 3 onTrial 2, then the full set (levels \-A) on Trials3 and 4. The thought here was that recall ofthe new level-4 words on Trial 3 might beappreciably aided by S having already hadpractice in recalling the Blocked level 1-3superordinate words. No such benefit wouldbe expected for the Progressive parts Ssreceiving the Random list.

MethodDesign and procedure. There weresix groups ofSs,

four learning by Progressive parts and two by theWhole presentation methods. Each S had four input-output trials on a list composed eventually of fourcompleteconceptualhierarchiescomprising 1 1 2 words.Because of the much shorter times requiredfor Trials 1and 2 in the Progressive conditions, we were able torun a complete replication of the experiment withthese Ss within the experimentalhour. The replicationsimply maintained the same progressive-parts condi-tions but with a different set of four hierarchies.The Ss were run individually. A complete or partialhierarchy was shown printed on a large 5 x 8-in. cardwith a total study time calculated at two seconds perword onthe card.The wordson each cardwerearrayedin the formofa vertical tree but without the connectinglines and circles shown in Fig. 1. The same tree-form

was used even for the Random lists. After seeing foursuch cards, S recalled thewords orally in any order hepreferred. Total time allowedfor recall was five secondsper word on the input list. No S everneededthis muchtime, and S typically initiated the next input trial byindicating that he could recall no more.

The six experimental conditions were as follows:(a) Whole Blocked (WB): The four complete hier-archies (averaging 28 words per hierarchy) werepresented in conceptually blocked

fashion,

one percard, for all four

trials,

(b) Whole Random (WR):The same set of 112 words, approximately 28 on eachof four cards, were presentedon each trial. The nodesin the tree on each card werefilledby randomselectionwithout replacement from the 112 words, avoidingobvious conceptual relations amongst words insuccessive nodesofa tree, (c) ProgressiveBlocked(PB):On Trial 1, each exposure card contained only thelevels 1 and 2words(14 intotal) inconceptuallyblockedfashion. On Trial 2, the appropriatelevel-3 words wereadded below the levels 1 and 2 words on each card(40 in total); and on Trials 3 and 4, the 72 level-4words wereaddedappropriately. Thus on Trials 3 and4, the exposure conditions for PB 5s were identicalto those of WB Ss. (d) Progressive Random 1 (PR1):This used the same trees as the WR condition, butwith progressive exposure of levels 1 and 2 words onTrial 1, levels 1, 2, and 3 on Trial 2, then the full4-level hierarchies on Trials 3 and 4. (e) ProgressiveRandom 2 (PR2): This was similar to PRI except therandomizationof words was within, not across, levelsofthe conceptual trees. A level-« wordin a conceptualtree was used only at a level-/! node in a PR2 tree; thedifferencebetween such random trees and the concep-tual trees is only in the vertical relationships amongwords above and below the nodal words. Thus thePR2 word sets being recalled over trials were identicalto those of condition PB, except that the scramblingof words across cards did not allow the PR2 Ss toeasily discover the conceptual relationships. (/) Pro-gressive Unrelated (PU): These Ss learned 112 un-related nouns by the progressive parts method, with14 on Trial 1, 40 on Trial 2, and all 1 12 on Trials 3and 4. The words were arrayed in vertical tree formas in the other cases. This condition was run forcomparison tothe PR conditionswhich had conceptualword lists. The 1 12 unrelated nouns werecomparablein Thorndike-Lorge frequency to the words of theconceptual list, but because they were all nouns, tendedto have higher "concreteness" ratings (Paivio, Yuilleand Madigan, 1968) than did the words on theconceptual list.

As indicated

before,

there was time to completefour trials on a second list of four differenthierarchieswith Ss in the Progressive conditions. The two sets offour hierarchieswereanimals, clothing, transportation,and occupations in set 1, and plants, instruments

326 BOWER ET AL.

>

i

tt

i

body parts, and minerals in set 2. Half of the Ss ineach progressive condition had set 1 first and set 2second, while theremaining Ss had the reverse order.Similarly, four of the eight Ss in the WB and WRconditions learned set 1 and four learned set 2.

The Ss were 48 Stanford undergraduates fulfillinga service requirementfor their introductory psychologycourse. There were eight Ss in each of the six experi-mental conditions.

ResultsFor the four Progressive groups, there was

no significant different in performance com-paring their List 1 with List 2, so theirperformance on both lists was pooled toincrease reliability of the following analyses.

TABLE 1Average Words RecalledoverFour Trials

Trials

Condition 12 3 4

Whole:Words presented 112 112 112 112

Blocked 73.0 106.1 112 11273.0 106.1 112 112Random 20.6 38.9 52.8 70.1

Progressive:Words presented 14 38.5 112 112

Blocked 13.0 34.5 93.2 10813.0 34.5 93.2 108.4Random I 8.6 16.0 34.9 46.3Random 2 8.4 15.5 33.9 52.6

The averagerecall scores for thefour trialsfor the six conditions are shown in Table 1.Looking first at the two Whole-presentationgroups, recall in condition WB is seen to bemarkedly superior to that ofgroupWR. Meanrecall is 3.5 times better in condition WB onTrial 1, and there is no overlap among therecall scores of the two groups of Ss on anytrial. Recall of the 1 12 words in condition WBis almost perfect by Trial 2. Obviously, thestructural organization of the blocked inputlist had a tremendously powerfuleffect on freerecall in this situation.

Turning next to the four Progressive condi-tions in the lower portion of Table 1, recall in

the Blocked condition is seenagain to be muchhigher than in the comparable Randomconditions. Recall scores in the two Randomconditions, PRI and PR2, do not differsignificantly from each other, nor do theydiffer significantly from recall of Ss learningthe unrelated word lists. All three of theseconditions are significantly inferior to condi-tion PB from Trial 2 onwards.

Performance on Trials 3 and 4 providescomparisons between the Whole versusProgressive parts methods, since on thesetrials Ss were exposed to the same full list.Within the Blocked conditions, WB signifi-cantly exceededPB on Trial 3 simply becausethere was no variance for WB Ss on this trial(all recalled perfectly). Within the Randomconditions, recall in WR exceeds that ofeither PRI or PR2 on both Trials 3 and 4.Thus, in terms of overall recall in thesecondi-tions, it was clearly advantageous to presentall words on all trials rather than progressivelyincreasing the exposure set over trials.

One may ask the further question whetherprior practice at recalling the level 1-3 wordssubstantially improves S's ability to recall thelevel-4 words when they are presented for thefirst time on Trials 3 and 4. The level-4 recalldata are presented for the relevant groups inTable2.

Level-4 recall for group PB on Trials 3 and4 significantlyexceeds level-4 recall for group

TABLE 2Recall

of

the Level-Four Words

Trials

Condition 12 3 4

Whole:Words presented 73.5 73.5 73.5 73.5

Blocked 44.6 69.6 73.5 73.544.6 69.6 73.5 73.5Random 13.2 22.8 35.2 44.8

Progressive:Words presented 0 0 73.5 73.5

Blocked 56.6 70.456.6 70.4Random 2 — 16.3 30.0Unrelated 13.8 27.6

Unrelated 9.9 18.7 32.4 51.1

327

HIERARCHICAL

RETRIEVAL

SCHEMES

WB on Trials 1 and 2 (t = 2.17, p < .05), sothat proactive facilitation exists when the newwords to be recalled fit as subordinates intothe conceptual structure practiced on Trials 1and 2. However, a similar facilitation appearscomparing Trials 3 and 4 of group PR2 withTrials 1 and 2 of group WR; a t test for thelevel-4 recall over the respective trials for thetwo groups is significant, ?(14) = 2.63, p < .05.For the unrelated words, comparing Trials 3and 4 of group PU with Trials 1 and 2 ofgroup WR, the difference in recall is in thesame direction but not significant. From theseseveral comparisons, the overall conclusion isthatrecall of a new subset of words presentedfor thefirst time within a larger set is facilitatedif there has been prior practice at recallingthe other words. This result may be specificonlyto thecomplete presentation method usedhere, since that permits S to spend relativelymore time studying the new words added tohis input list.

Further analyses of recall protocols demon-strated that WB Ss were essentially using thestructural information in the blocked inputas a retrieval plan for generating their recall.First, in condition WB, Ss' recall was organ-ized almost entirely according to conceptualcategories, whereas Ss in conditionWR tendedto show conflict between clustering accordingto the underlyingconceptualcategories versusaccording to the spatially contiguous wordsin the randomized input. Second, all WB Ssrecalled their hierarchies from level- 1 downthrough level-4 words (from top down), andneverin the reverseorder, in doing this, someSs in condition WB generated the conceptualtrees in breadth across levels (i.e., all thelevel-« words of a hierarchy before any of itslevel-rt+l words), whereas other Ss generatedthe hierarchy in depth, along a given (say)left-going branch before returning to highernodes to generate in depth the other branchesfrom it. Most WB and PB Ss claimed that theycould have reproduced the hierarchical arraysexactly as shown had they been permitted towrite their recall ; giventheir perfect oral recall

and the facts above, this is a credibleclaim.

A question related to the results above iswhether recall of a level-« word for WB Ssserves a cuing function, being correlated withrecall of the level-«+l instances nested withinit. To answer this question, weconsidered thelevel-4 recall conditional upon level-3 recallby WB Ss on Trial 1 where there were the mostdata and recall failures. The average prob-ability ofrecalling a level-4 word was .66 whenits level-3 word was recalled, but only .30 whenits level-3 word was not recalled. Moreover,when a level-3 word and at least one of itslevel-4 words were recalled, in 90% of thesecasesrecall of the level-3 word preceded recallof the first level-4 word (not necessarily inimmediate succession, however). Those caseswhere level-4 words were recalled withoutprior recall of the level-3 word were most ofteninstances in which the level-4words werecued(preceded) by a level-2 word, skipping thelevel-3 word. These results accord with theexpectation that recall of a word served to cuerecall of the corresponding words at lowerlevels. When a nodal word was not recalled,the entire "tree" developing out of that nodewas likely to be missing in recall.

In contrast to this strong dependencebetween recalls of conceptual levels 3 and 4for the blocked hierarchies,recall of thesesamewords was virtually independent for Ss whoviewed random hierarchies. Scoring Trial 1recall of WR Ss for the same words as above,the conditional probability of recalling alevel-4 word was .23 when its level-3 word wasrecalled, and was . 1 7 when its level-3 word wasnot recalled. These probabilities reveal nocontingencies in excess of chance expectationon the null hypothesis of independent events,X2(l)=Ls, p>.3o. Thus, the influence onrecall of the moderately strong normativeassociations from a level-« category word to alevel-/7+l instance was considerably modu-lated by the method of presentation. If theassociated words were presented in closetemporal and spatial contiguity, so that S

328

BOWER

ET AL.

might rehease them together, they tended tobe recalled together; if they appeared widelyseparated in a random list, then the samewords were recalled almost independently inthis case.

There were some small differences in therecallability of the eight conceptual hier-archies. Only the WB and WR conditions onTrials 1 and 2 were considered for thesecomparisons since, by the time the full hier-archies were presented, thePB Ss were makingvery few errors. The proportions of wordsrecalled for the various hierarchies averagedover Trials 1 and 2 for WB (and WR) Ss wereas follows: Plants, 87(26); minerals, 85(21);body parts, 84(34); transportation, 81(16);instruments, 79(27); occupations, 77(25);animals, 76(24); clothing, 72(29). The rangeof variation in recall of the word sets is about15% for both groups, but therank ordering ofthe sets by the two groups is quite discrepant(rank correlation =—.14). The poor correla-tion of the recall ranking of the eight hier-archies in conditions WB and WR may havebeen due to treatment effects or to lowreliability since only four Ss in a givencondition learned a given hierarchy.

Perhaps the main fact to be rememberedfrom Exp. 1 is that complete presentation ofthese conceptual hierarchies produced atremendous facilitation of free recall relativeto a random ordering of the same words. Theeffect is similarto what one finds in free recallof a sentence versus scrambled words (e.g.,Miller and Selfridge, 1950). In both cases,recall depends upon the way the words arearranged, with afamiliar structure in one casecontrasted to unfathomable randomness inthe other.

The notion of hierarchical organizationas arecall aid is hardly original with us. Mandler(1967, 1968) has hypothesized that recall islargely a matter of subsuming list items undera hierarchical array of categories. Someearlier experimental work was done by Cohenand Bousfield (1956) who were interested inrecall clustering produced by either (a) a

40-word list classifiable into eight independentcategories of five instances each, versus(b) a dual-level list consisting of four super-ordinate categories each divided into twosubordinate categories (e.g.,feline animals andcanine animals), with five instances in eachof these eight subordinate categories. Cohenand Bousfield reported that total recall ofthese two lists was very similar (17.6 vs. 18.1out of 40), but that recall clustering differedsomewhat. The Cohen and Bousfield studydiffered in a number of respects from ours.For example, in their study thecategorynameswere neither presented nor recalled, and itemswere presented singly in random order. Theseprocedures surely would reduce the probabi-lity that S would notice and utilize the dual-level organization of the second list. In thisregard, we may point out that our PR Ss,who received a hierarchical word list butrandomized, did not recall any better than ourPU Ss who had unrelated words. Obviously,the list-structure has to be discovered andutilized if S is to derive any benefit from it inhis recall.

Complete presentation of the Blockedhierarchies provides S with a lot of structuralinformation about the word list—he doesnothave to discoverit for himself. This structuralinformation in turn provides S with a planfor retrieving the words from memory. Thesalient characteristics of a retrieval plan arethat it tells S where to begin his recall, how toproceed systematically from one unit to thenext and to the words within each unit. Theplan also helps S to monitor the adequacy ofhis recall, helping him to identify where partsare missing and to identify when he hasfinished. It is plausible thatacentral ingredientin the present hierarchical retrieval plan isassociative cuing of the words at level-«+l byrecall of thesuperordinate categoryat level-n.Having alreadyrecalled masonry (stones), S isset to search for recency-tags on words in hisassociative hierarchy to this category (e.g.,granite, limestone, sandstone, flagstone, etc.).It is further plausible to assume that the

329HIERARCHICAL RETRIEVAL SCHEMES

\

implicit candidates generated to the categorycue are recalled only if theypass a recognitioncriterion as having been on the list. Thus, ofthe four masonry-stone candidates, S mightovertly recall granite and limestone but notintrude sandstone andflagstone since the lattertwo words werenot tagged as having been seenrecently.

The detailed analyses of therecall protocolssupport this general view. When recalled,the superordinate category almost alwayspreceded recall of its instances, and instancerecall was considerably poorer if the super-ordinate was not recalled. Moreover, theintrusion data in group WB accord with theexpectationfrom this category-cuing analysis.There were relatively few intrusions (20) onTrials 1 and 2 for group WB, but all wereobviously intrusions of unpresented instancesof a presented category, and these intrusionswere "appropriately" placed in the recallorder. Group WR had 26 intrusions ofwhich25 wereclearly within thepresented categories.In theory, these intrusions represent simple"false alarms" in recognition of a likelycandidate which in fact was not on the list.The theory supposes that intrusions decreaseand correct recalls increase with practicebecause of two factors: (a) The associationsfrom category to presented instances becomestronger, thus causing the latter to be morereadily and reliably generated as candidatesfor recall to the former, and (b) the item-information that aids list discriminationimproves, which information may be in theform either of an estimate of frequency orrecency of experience or a trace-strength ofthe presented items.

Experiment IIThe next experiment investigated whether

the large influence of Blocked structural-information upon recall could be replicatedwith arecognition test of memory. A frequentclaim in the recent literature on memory isthat recognition measures the amount ofinformation stored independently of retrieval

processes. In terms of the theory outlinedabove, the occurrence ofan item on the inputlist causes information of some kind (a tagdenotingrecency, frequency, or trace strength)to be attached to the representation of thatword in S's semantic memory. In a test forrecognition (or discrimination of list member-ship), the occurrence-information attached tothe test-word in memory is consulted for adecision, yielding a judgment of recognitiononly if theinformation thereexceedsa criterion(e.g., Bower, 1967; Parks, 1966). Suchrecognition tests, which directly provide thetest word, clearly bypass the search andretrieval processes by which S generates hisrecall. If the structural information providedby the blocked hierarchies has its maininfluence on the retrieval plans for recall,then one should find much less of an effect ofstructural information when memory is testedby recognition, which largely bypasses theretrieval aspects of the task. Accordingly,Exp. II was undertaken to see whether thelarge recall difference between the WB andWR conditions would be greatly attenuatedin recognition tests.

MethodDesign andprocedure. Each S received two replica-

tions of a two-trial experiment, using the four-hierarchy sets 1 and 2 from Exp. I. There were eightconditions obtained by crossing Blocked vs. Randompresentationwith Recall vs. Recognition over the two

trials of each learning task. Using the last-letterabbreviations L and N for recall and recognition,respectively, the four possible test sequences over twotrialsareLL, LN, NL, and NN. Subjects in conditionsLN and NL were switched between Lists 1 and 2, as

were Ss in conditions LL and NN.The fourhierarchies of set 1 or set 2 wereprojected

on fourslides in tree

form,

at56 secper slide, and thenSeither wrote his recall or took a written recognitiontest. He did not know during presentation how hewould be tested. The recognition test used the "Yes-No" method: A sheet of paper given to S containedthe 112 list words scrambledamongst 112 distractors,and S was told to check those words which he thoughthad been on the slides just studied. Half of the112 distractors were unrelated nouns matched inThorndike-Lorge frequency to the list words, whilethe other half of the distractors were conceptually

330 BOWER ET AL.

related words that could have fit into the conceptualhierarchies presented. The same set of words wereused for both recognition tests in the NN condition,except they were listed in a differentorder on the twotrials. Subjects were run in small groups of

two-four,

with test times sufficient for all Ss to finish. A distinctinstructional break was made between Lists 1 and 2.Half the Ss were exposed to the Blocked lists and halfto the Random lists of Exp. I for both lists and bothtrials. The Ss were 64 Stanford undergraduates fromthe introductory psychology course with eight in eachof the eight groups. Pooling across the two lists, thereare 16 learning protocols for the eight experimentalconditions.

Results

There were no significant differences be-tweenfirst versus second lists nor between sets

TABLE 3Percentage

of

Words Recalled in the

Six Conditions

Condition T,

Blocked:L LL N'N L

.61

.65

Random:L LL NN L

.21.29

° L N denotes recall on T, and recognition on T2 .1 and 2 of the hierarchies, so the data will bepooled over these variables. The first resultof interest is shown in Table 3 which givespercentage correct recalls over Trials 1 and 2for the Ss that had recall tests. The recalldifferences between Blocked and Random Ssare of similar magnitude to the differencesobserved in Exp. I, so that result is replicable.Further, recall on Trial 2 was approximatelythe same whether Trial 1 was a recall or arecognition test. Recall intrusions were lessthan 1 % in all six conditions, but werecategorical in nature. The Blocked Ss gave63 intrusions of which all were categorical;

the Random Ss gave 31 of which 29 werecategorical.

Recall of the eight conceptual hierarchieswas separately scored, pooling all Ss who wererecalling them on either Trial 1 or 2. Therecall proportions for the eight hierarchiesaveraged over Trials 1 and 2 for the Blockedcondition (Random condition in parentheses)were as follows: Body parts, 81(34); plants,75(35); minerals, 71(34); animals, 71(33);clothing, 71(35); transportation, 69(26);instruments, 69(32); and occupations, 61(34).The variation in recall of seven of the eightrandom hierarchies is too minuscule to givemuch of any correlation with recall of theeight blocked hierarchies (/- =+. 17). Thesescores may be compared to those obtainedin Exp. I over the same trials but with fewer Sslearning each hierarchy there (four Ss insteadof 16). For blocked presentation, rank ordersof the recallability of the eight hierarchies inthe two experiments correlate +.62; forrandom presentation, the two rank orderscorrelate +.48. It thus appears that thedifference in recallability of the eight hier-archies is small but reasonably consistentwithin a givencondition, but that the orderingdiffers consistently for blocked vs. randompresentations. We shall look into thisissue again in Exp. V in which severalnormative indices of the hierarchies areobtained.

TABLE 4Recognition Hits (H) and False Alarms

on Related (R) and Unrelated (UR) Distractors

Trial 1 Trial 2

Condition H R UR H R UR

Blocked:L N .92 .04 0N L .88 .05 .01N N .80 .07 0 .92 .08 0

Random:L N .78 .06 .01N L .61 .06 .03N N .60 .10 .02 .80 .19 .07

331HIERARCHICAL RETRIEVAL SCHEMES

The relevant results on recognition testsare shown in Table 4 for Trials 1 and 2. Thethree columns for each trial give proportionsof (a) checking as "old" an old item (a hit),(b) falsely checking a conceptually relateddistractor,and(c) falselychecking an unrelateddistractor. Hit proportions are based on 112observations per S per trial, whereas the falsealarm proportions are each based on 56observations per S per trial.

Comparing entries within Table 4, thefollowing conclusions are warranted; (a) Falsealarms were higher on related than onunrelated distractors; (b) hit rate on Trial 2was unaffected by type of test on Trial 1,but the false alarmrate on Trial 2 was higherif Trial 1 was also a recognition test; and(c) Ss having the Blocked input have betterlist discrimination than Ss having Randominput, as indicated by a higher hit rate anda lower false alarmrate.

The first conclusion, that related distractorsare checked more often, is hardly surprisingand can be handled theoretically in severalways. The second conclusion, that false alarmsare highest in the NN conditions, is probablyaresult of using the same distractorson Tests 1and 2. Thus, on Test 2 in the NN condition,a distractor arouses a familiar sense of havingbeen seen before and is checked more often.These are ancillary findings.

The interesting result is the third, thatrecognition is better for Blocked than forRandom Ss. A t test on recognition hitsminus false alarmsfor all Blocked vs. RandomSs on Trial 1 yields r(62) = 5.94, /x.OOI,and on Trial 2 yields ?(62) = 6.69, p < .001.Comparing the averagedBlocked vs. Randomresults in Table 4 to those in Table 3 showsthat the averagerecall differences were muchlarger than the recognition differences; recalldifferenceswere 38 and 48 % on Trials 1 and 2compared to recognition differences of 23%and 13% on the same trials.

We may ask whether the recognition testhas attenuated the difference betweenBlockedand Random Ss seen on the recall test.

However, an interpretative problem is thatthere is no atheoretical way to evaluate thesignificance ofthese differences due toBlockedvs. Random conditions in the recall vs. recog-nition measures. A model relating free recallto recognition is needed, but none can bestated now with any confidence. A simplethreshold model, which supposes thatS recog-nizes all those items he could recall plus halfof those he could not recall, leads to theformula 2N=\+L, and this gives a fairlygoodprediction of the average hit rates (Table4) from the average recall results (Table 3).However, it gives no prediction offalse alarmrates, nor their ordering, nor does it provideany illuminating analysis of the retrievalprocesses in free recall.

The results of this experiment suggest thatrecognition of list membership of a worddepends on the structure of the list as well asthe words that were in it. In particular,recognition of a given word apparentlydepends upon the number and configurationof associations converging upon that wordfrom other list words that have been recentlyprimed. This view accounts for the effect oflist structure upon hit rate, and it also explainswhy related distractors elicit more falserecognitions than unrelated distractors, evenin the Random condition. Underwood (1965)and Anisfield andKnapp ( 1968)have proposeda similar view ofrecognition memory.

An alternative interpretation of theserecognition differences would attribute theeffect to "unequal" exposure times to theindividual words in the Blocked vs. Randomhierarchies. Since only total time was con-trolled, Ss in the Blocked conditions mighthave used the redundancy or predictabilityof the hierarchies to scan rapidly and moreoften over all the words. On the other hand,Ss in the Random condition may have beenmuch slower in reading the words, and fortworeasons: (a) Recognition or reading timesof words are known to be slower if the wordsequence is unpredictable, so fewer wordswould be scanned in a fixed time, and (b) the

332

BOWER

ET AL.

Random S might have used up exposure timetrying to organize or systematize the randomwords he had already read, thus reducing thetime he could spend processing later words.According to this view, then, the Blocked vs.Random difference in recognition resultedfrom lessreal exposure to the individual wordsin the latter condition. We are currentlytesting this interpretation with a method ofpresenting the hierarchy one word at a time,thus permitting control of S's exposure timeto each word. The S sees a complete hier-archical tree of unfilled, nodal circles withlines between nodes, and 28 successive slidesshow in systematic order one new nodal wordfor 2 sec, until each word in the conceptualhierarchy has been presented once. If differen-tial exposure time to the individual words werethe factor determining the results of Exp. 11,then the Blocked vs. Random difference inrecognition should vanish whenexposure timeis controlled in this manner.

Experiment 111The next experiment investigates the retro-

active effect upon free recall of a first listinduced by the learning of a second list.Previous experiments by Tulving and Thorn-ton (1959), and Postman and Keppel (1967)have shown retroactive decrements in freerecall. For example, in the Postman andKeppel study, Ss learned List 1, then hadvarying numbers of trials on List 2, thenwereasked to recall all the words in both lists.Recall of List-1 items was markedly poorerthe greater the number of trials S had had on

List 2. The authors discussed these results interms of the unlearning (during trials on

List 2) of the associations between the experi-mental context stimuli and List-1 responses.

These interferenceor unlearningeffects havebeen produced with unrelated words in Lists 1and 2. Our question here is whether we canproduce just the opposite effect, retroactivefacilitation, if the two word lists fit into thesame conceptual hierarchies. In particular,retroactive facilitation of List-1 recall might

occur if List 1 consists of level-4 instances ofconceptual hierarchies, and List 2 consists ofthe levels 1-3 superordinates of these samehierarchies. In this case, when S is then askedto recall everything from Lists 1 and 2, hisrecall of the superordinates from List 2 shouldserve a cuingfunction in facilitatinghis recallof the List-1 instances.

MethodDesign andprocedure. There were threeexperimental

conditions, a rest control (C), relevant interpolation(Rl), and irrelevant interpolation (II). All Ss beganwith two trials on List 1 which consistedof48 level-4words from two conceptualhierarchies (e.g., mineralsand animals). These hierarchies were so tailored thatthey each had six groups of four level-4 words. Thesefour level-4 words (subsumed by one level-3 node)werepresented together asa group for eight seconds bya slide projector. In List 1, there were 12 such slidesarranged in a differentrandomorder forthe two trials.Subjects gave theirfree recall by writing the list words,being permitted 3 min.

Aftertwo trials on List 1, the Rl and II Ss learned asecond list while the C Ss read Peanuts cartoons for atime equal to the longest interpolatedlearning interval.The interpolated list differed for Rl and II Ss. For RTSs, theinterpolatedlist consisted of the levels 1 , 2 and 3superordinate words from the two hierarchiesrelevantforclassifying the level-4 wordswhich S had learned inhis List 1 (e.g., mineralsand animals). The relevance ofList-2 to S's List-1 words was not pointed out to him.For II Ss, the interpolated list consisted of the levels1, 2 and 3 superordinatewords from two hierarchieswhich were irrelevant for classifying the prior List-1words (e.g., instruments and occupations.) The twoList-2 hierarchiesfor a given S were presented in tree

form,

one per slide for 18 sec (two seconds per ninewords ineach tree), andthenS wrotehis recall ofList 2.

After S reached a criterion of oneperfect recall ofList 2, hewas thenaskedto recall(in writing)everythinghe could from both Lists 1 and 2, starting with List 2if he could. Rest control Ss were also asked to recallList 1 at this time.Four minutes wereallowed for thisrecall and protocols werecollected. At this point, Ssin groups C and IIweregiven a sheet of papercontain-ing the relevant superordinate categories for List 1(i.e., levels 1-3 arrayed in tree form) and were askedto attempt a second recall, of the List-1 words only,since "the words on this sheet may help you to dobetter in your recall" of List 1. Three minutes wereallowed for this cued recall by Cand II Ss.

After the procedures described above were com-pleted, the entire experiment was replicated with

333HIERARCHICAL retrieval schemes

completely different word hierarchies partitioned toserve as List 1 and List 2 but with S assigned to thesame treatment condition as before. The order inwhich the various word hierarchies were used (first orsecondreplication) was counter-balanced over Ss.

The Ss were 33 undergraduates fulfilling a servicerequirement for an introductory psychology course.They were assigned 10 to group

C,

11 to group Rl,and 12 to group 11. The Ss in a given condition wererun in small groups of two to five at a time.

ResultsThere was a significant improvement in

Ss' performance from replication- 1 to repli-cation^. Considering List-1 performancewithin each replication, recall proportionsaveraged 52 and 68% on Trials 1 and 2 ofreplication- 1 , but were 59 and 79 %on Trials 1and 2 of replication-2. This general practice

TABLE 5

Recall

of

the 48 Level-Four Words

of

List 1

List 2Rest

Control Irrelevant Relevant

1. Trial 2 of List 1 35.3 34.5 35.62. Recall all 34.9 34.5 42.23. Relevant probe 40.3 40.1

effect, however, is orthogonal to the majorfactors of interest in this experiment, soresults will be pooled over replications for theanalyseswhich follow.

There were no significantdifferences amongthe three groups in their acquisition of List 1.The average List-1 recall scores on Trial 2(the last acquisition trial) are shown in thefirst row of Table 5 and are obviously similar.Learning of the List-2 blocked hierarchies byRl and IISs was veryrapid, with the median Srecalling his List-2 perfectly onTrial 2. TheseSs also recalled their List-2 perfectly on the"recall everything" test which immediatelyfollowed their criterion trial.

The results of major interest are the List-1recall scores for the "recall everything" testgiven after List-2 learning; these are recordedin the second row of Table 5. Comparingthis

recall to the corresponding recall scores onTrial 2 (row 1), we see that there was practi-cally no change for the C and II Ss, but a largeincrease for Rl Ss. The increase in recall forRl Ss is highly significant: r(21) = 7.95,p < .001. On this later test, the gain in recallrelative to what Rl Ss could have gainedwas 53%.

A similarly significant increase in List-1recall occurred for groups C and II when therelevant superordinates hierarchies were givenas recall probes on their last test (row 3 ofTable 5). The II Ss gained 42% and the C Ssgained 39% of the responses they could havegained over Trial 2 with List 1. A comparisonof the probe recall scores of C and II Ss withthat of Rl Ss on their "recall everything" trialyielded no significant differences, F(2, 28) =

2.07,p>. 10.The simplest explanation of these results

would attribute the List-1 recall increments tothe cuing function of the relevant super-ordinate words. That is, onTrial 2 with List 1,wholegroups of level-4 words might not havebeenrecalled because S had not yet developedascheme for cuing his recallof these. However,when the relevant superordinate categoriesare available on S's recall sheet, either by E'sprovision or by S recalling them, they leadhim to think about practically all the groupsof level-4 words and to recall some words ofthese groups. This account is bolstered byanalysis of the cluster-recall data. We scoreda level-4 cluster as recalled if at least one ofthe four words in that group was recalled.For the clusters that were recalled, we alsocomputed the mean words recalled per cluster.These measures for Trial 2 ofList 1 andfor thefinal recall trial are shown in Table 6 for thethree treatment groups. The groups differrelatively little in the percentage of clustersrecalled or in the mean words per recalledcluster. The effect of hierarchical cuingprovided by S or E was to increase thepercentage of clusters recalled but not themean words per recalled cluster. Within eachgroup separately, the increase in proportion

334 BOWER ET AL.

TABLE 6Proportion

of

Clusters Recalled and MeanWords per Recalled Cluster on Trial 2 and Final

Recall Trial

Proportion of Words per clusterclusters (of4)

Group Trial 2 Final Trial 2 Final

Relevant .84 .98 3.52 3.58Irrelevant .89 .99 3.20 3.35Rest Control .86 .98 3.40 3.42

of clusters recalled is significant (/ = 8.00,12.20, 5.80 for groups Rl, II and C, respec-tively); but for no group was therea significantchange in the mean words perrecalled cluster.

We have attributed the retroactive facilita-tion of recall in group Rl to the simple cuingeffect of having the superordinates availableat the time of recall. This is in accord with theresults of Tulving and Pearlstone (1966) andWood (1967b), who found more wordsrecalled if Ss were provided with relevantcategorycues at the recall test. An alternativeinterpretation might attribute the Rl resultto strengthening of List-1 words while S wasstudying the relevant List-2 superordinates.That is, while viewing the relevant super-ordinates, S might notice their relationship toList-1 responses and implicitly rehearse theList-1 responses, associating them with thesuperordinate context of List 2. This viewwould provide a "mediational" interpretationof our Rl result similar to that provided forthe retroactive facilitation observed with theA-B, A-B' paradigm in paired-associateslearning, where A is the cue, and B and B'are associatively related response terms (cf.Postman, 1961). An implication of this inter-pretation, however, is that final recall for RlSs should have exceeded probed recall for theC and II Ss, since the Rl Ss allegedly have hadthe benefit of extra rehearsal on List-1 wordsduring their List-2 learning. In fact, however,there was no reliable difference between finalrecall for Rl Ss and probed recall for C and IT

Ss. Thus, this one implication of the media-tional interpretation was not supported by theevidence.

Experiment IVThe first three experiments have examined

theinfluence ofstructured input upon memoryfor conceptual hierarchies. The relationshipbetween successive nodal words in a con-ceptual treeis roughly one of "class inclusion,"or "instance of a superordinate category."In terms of associationistic psychology,however, these class inclusion relationshipsare only one of many bases upon which twowords could become associated. Taking thisbroader view, then, a conceptual hierarchy isjust a specific kind of associative hierarchy ornetwork, one for which the associationsbetween nodal words have approximately thesame basis, namely, class inclusion. But thisbroader view immediately leads one to con-sider other kinds of associative trees—specifically, trees in which successive nodesare associated with one another,but in whichthe basis for the association may vary freelyover different nodal pairs in the same tree.

The tree in Fig. 2, for cheese, will illustrateour points about associative hierarchies. Weconstructedsuch trees by startingwith a level- 1nodal word, and recording three different butstrong verbal associates to it as level-2 nodes.Each of these second-level nodes was thenconsidered in turn and to each was recordedtwo different associates as level-3 nodes;finally two associates were recorded beloweach level-3 node, giving a four-level hier-archy of 22 words. We have constructedeight of theseassociative trees. They are basedonly on intuitive judgments of pairwiseassociations, since published associationnorms are too restricted to provide guidelinesfor constructing such depth hierarchies. Thereis obviously nothing unique or special aboutsuch associative hierarchies. There are doubt-less thousands that could be constructedstarting with almost any content word innode-1, with variation in the number and

HIERARCHICAL RETRIEVAL SCHEMES 335

Fig. 2. The associative hierarchy for "cheese."

depth of branchings. The associative base forthe hierarchies in Fig. 2 is not easily specifiedsince the type of association relating node-/?to node-fl+l varies considerably. Associa-tions between successive pairs are "intuitivelysensible," but, like free-associative chains, theconnections among more distant elements aresensible only by virtue of the interveningpairwise associations in the chain.

Given theseassociative hierarchies, one canagain ask the same question as in Exp. I;namely, is free recall facilitated by presentingthem in a structurally blocked fashion asopposed to a random fashion? Accordingly,a simple comparison was carried out betweencomplete presentation of these blocked hier-archies and complete presentation of thesame words mixed up randomly. The twoconditions are similar to the WB and WRconditions of Exp. I.

MethodThe Ss were 44 students in a high-school plane

geometry class made available to us in a local publicschoolsystem. Thestudents were mainly tenth graders,with a few from grade 1 1. They were tested for 30 minon two consecutive days, having four input-outputtrials on differenthierarchies on each ofthe two days.Their regular teacher cooperated as one ofthe £s andin proctoring their performance.2

Design and procedure. The Ss received mimeo-graphed booklets of the study materials and recallsheets. Each booklet consisted of 12 pages, four setsofthree pages, consisting of two study pages with list

2 Our thanks to Mrs. Sharon Lesgold, the teacher,for helping us with this experiment.

words followed by a blank recall sheet. Instructionswere given orally by E.

There were four types of booklets distributedeachday,each to one-fourthof theclass. One type presentedtwo hierarchies in an organized (blocked) tree

form,

and the second used these same words in two trees,but with total scrambling of the words across associa-tive hierarchies and levels. The thirdand fourth typeswere the same as the first and second types, exceptthat the order of presentation of the two hierarchieson each trial was reversed. On the first day, the wishand hammer hierarchieswere used, and on the secondday the salt and cheese hierarchies were used. Thus,the second day was a replication of the

first,

with newwords, except that each S received the conditionopposite to what he had received on Day 1 (i.e.,organized-random order or vice-versa).

The hierarchies of22 words wereprinted in verticaltreeform(with no circles or lines), oneper study sheet.Each was studied for 45 sec as timed by a stop-watchby E, with instructions to turn the page to the nextstudy sheet. After studying both hierarchies, at asignal S turned to a blank recall sheet and tried to writeall the words he had studied, in any order. Recall timewas 34- min; its termination was indicated by Eand the next study trial began.

One S was dropped because he was not availablethe second day, and one because she obviouslymisunderstood the instructions on the first day. Thisleft 42 usable Ss with two learning protocols each.

ResultsThere wereinsignificant differences between

recall on Days 1 and 2 within a given condi-tion, so all protocols within acondition werepooled ignoring this Days factor. The meanwords recalled (out of 44) for the Blockedcondition on Trials 1-4were 23, 34, 39, and 41 ;the comparable means for the Random

336 BOWER ET AL.

condition were 16, 23, 28, and 33. Meanrecall in the Blocked condition significantlyexceeded that of the Random condition oneach trial (all p's < .01). Each S learned botha Blocked and a Random list; 40 of the 42 Ssrecalled better with their Blocked list.

Recall of the four different associativehierarchies was quite homogeneous. Percent-ages recalled averaged over the four trialswere as follows for the Blocked (and Random)hierarchies: Wish, 81(61); hammer, 79(59);cheese, 77(59) ; andsalt, 74(56). The differencesin recallability of thefour hierarchies are smalland insignificant within both the Blocked andRandomconditions.

An analysisof ordered recall similar to thatof Exp. I (which used oral recall) was not donebecause of equivocation in deciding in whichorder many Ss had written down the wordsthey recalled. Some Ss in the Blocked condi-tion wrote their recall in the tree form theyhad studied; but the temporal order in whichthey had recalled the nodes could not bedetermined. It was possible to score theprotocols for jointrecall of the level-3 and itslevel-4 words. Considering Trial 1 with theBlocked list, the probability of recalling alevel-4 word was .86 when its level-3 word wasrecalled, but only .36 when its level-3 wordwas not recalled. The comparable conditionalrecall proportions for the same words in theRandom lists were .39 and .32, respectively.A similar pattern held for recall of level-3words conditional uponrecall of level-2 words.This pattern of results again points to thecuing role of the level-« node in therecall of itslevel-«+l words. This cuing effect is muchmore potent with the Blocked than with theRandom hierarchies presumably because the«-to-«+l association is immediately availablefor rehearsal in the Blocked list, but is buriedamong many competing responses in theRandom list.

While the recall differences between theBlocked and Random conditions are largehere, they are not as dramatic as thedifferencesobtained with the conceptual hierarchies in

Exp. I. There were, ofcourse, many proceduraldifferences between the two experiments—different S-populations, different exposureand recall methods, etc. Ignoring theseprocedural variations,however, itappears thatthe blocked associative hierarchies are moredifficult than the blocked conceptual hier-archies, whereas the two random conditions(conceptual vs. associative words) wouldappear to be about comparable in difficulty.A directexperimental comparison of the twotypes of hierarchies is being done to checkon this ordering.

The main conclusion to be gathered fromthe results of this experiment is that blockingof associative hierarchies considerably im-proves recall over a condition of randominput of the same words. What is apparent toBlocked Ssbut not to Random Ss is that manyword pairs are associated and moreover thatmany lengthy associative chains exist in theword lists. The "structural principle" for theBlocked lists is that of recursive rewritingaccording to associative transitions, goingfrom each level-« node to its several level-n+lnodes. Subjects notice and comment upon thisconstruction principle, and theyobviously useit for reconstructing the list from memory.Again however, the recursive principle is notallowed to run unchecked, since it alonewould produce many associative intrusions ofnonlist words. Rather thecandidatesproducedby the associative recursion must be checkedfor list-membership information before theyare overtly recalled. The few intrusions thatdo occur for the Blocked Ss, however, appearpredominantly to be "false recognitions" ofcandidates generatedby the associative recur-sion principle.

Experiment VA possible objection to our experiments so

far might be that they merely show thatBlocked Ss have learned the principle of listconstruction, and that knowledge of thisprinciple is sufficient for them to produce

hierarchical retrieval schemes 337

most of the words in the hierarchies withouthaving any further specific information aboutthe input list. For example, Ss told to writethe numbers from 1 to 112 would producemore than Ss shown these same numbers onmany slides in a randomly scrambled orderand told to recall. The difference in "recall"would be utterly trivial in this instance becausethe list-construction principle suffices com-pletely to generateall list items and no nonlistitems. Are ourBlocked vs. Random compari-sons in the prior experiments simply showingthis same trivial outcome in a slightly veileddisguise?

We think not. The difference between thenumbers case and our hierarchies is in therange of acceptable alternatives that could begenerated by the list-construction principle inthe twocases. In the former case, the principleby itself generatesall list items and no others;in the latter case, the principle generatesonlya few list items and very many nonlist items(intrusions?). We shall return later to adiscussion of the psychological distinctionbetween these twocases ;but for our immediatepurpose it is best to collect some evidence forour claim that the word-hierarchies we haveused here are quite open-ended and non-exhaustive.

Experiment V essentially collected associa-tion norms on these hierarchies from a groupof Ss. The S was given a general descriptionof the list-construction principle and thentried to generate the word hierarchies. Hisgeneration was always set back on the "righttrack" when he erred. Given the completetree up through the level-« nodes, S was togenerate the appropriate number of level-M+lwords under each node. After this, the"right" level-«+l words were added to thetree, and S was asked to produce the appro-priate number of level-n+2 words. Both theconceptual and the associative hierarchies ofExps. I and IV were used. The question wassimply, how well will Ss do in generating suchtrees knowing only the principle by which thetrees were constructed?

MethodThe Ss, ten Stanford undergraduate volunteers

enrolled in Summer classes, were run individually.After general instruction illustrated with one example,S began work with the four associative hierarchiesfrom Exp. IV, and then did the eight conceptualhierarchies from Exp. I. The order of working on thehierarchies within each set was randomized over Ss.All work was completed on one hierarchy before Swent on to the next hierarchy. Work on a hierarchybegan by giving Sa slip ofpaper with the level-1 wordprinted in a box, and with 2 (or 3) lines radiating downto two (or three) empty boxes at level-2. The S wrotein what he thought were the appropriate words. Hewasthenshown a second slip ofpaper with the "right"level-1 and level-2 words, with lines radiating down tothe appropriatenumber of level-3 empty boxes whichhe was to fill. After filling these, he was given a paperwith the level 1-3 words and filled in the appropriatenumber of boxes for level-4 words. The recursive"resetting" of S back on the "right" track gave himfeedback on how well he was doing. Occasionally Scould not think ofenough appropriatewords to fill theindicated number ofempty boxes,and he was permittedto leave such spaces blank. Testing on the completeseries required 30-45 min depending on S's speedof generating appropriate words. Because of anexperimental error, responses of one S were lost onlevel-2 responses to one conceptual hierarchyand level-3 responses to another conceptual hier-archy.

Scoring. For notation, let k denote the total numberof responses required of S at a given level ofthe tree;k is the product of the number of nodes at that leveltimes the number of responses required at each node.With ten Ss, we shall have a distribution of 10A: re-sponses (minus a few omissions) at each level of eachhierarchy. A variety of indices could be computedfrom such associative distributions. We shall reporttwo.The firstmeasure will be theproportion ofthe \okresponses which match the "correct" words at thatlevel of the hierarchy. This will be relevant for inter-preting S's ability to generate the particular hier-archies which E had constructed.

The second measure is one of dispersion, or con-versely, how much Ss agreed on their responses ata givennode. A convenient measure ofagreement maybe obtained as follows: the responses at a given levelof a tree are rank orderedaccording to their frequency(with omissions considered as unique responses), andthen we see how far downin therank ordering one hasto go to accumulate 50% of theresponse distribution.The more agreement of Ss' responses, the lower willbe the rank of the median response. The rank of themedian will dependon k, unfortunately, so this scorewill not be comparable across levels unless it is trans-

338

BOWER

ET AL.

formed to a common scale. This is achieved in theAgreement Index(Al),

Nk2 X N 2x

Nk k 2V-1 k(N~l)2 2

where N is the number of

Ss,

k is the total responsesper S at a given level,and x is the obtainedrank ofthemedian word at that level. The Al index has thefollowing rationale: (a) If all Nk responses differ(i.e., no agreement), then the median rank,

x,

will beMfc/2 and Al will equal 0; and (b) if all N Ss agree ingiving the same k responses, then the median rank,

x,

will be kj2, and Al will equal 1.

Therefore,

Alscales from 1 for complete agreement down to 0 forcomplete disagreement.

ResultsThe values of these indices for the four

associative hierarchies of Exp. IV and for thesix conceptual hierarchies of Exp. I are shownin Table 7. The total proportion correct forthe overall hierarchyis obtained by a weighted

average of the proportions correct at eachlevel, with weighting coefficients equalling theproportions of words in the hierarchy contri-buted by the words at a givenlevel. Thecolumnlabeled NR reports the total number ofdifferent responses (types) givenfor theentirehierarchy by the 10 Ss, counting omissionsas uniqueresponses (i.e., considering them tohave been filled in by random selection froma dictionary). Excluding the level-1 name ofthe hierarchy which S was given at the start,the target associative hierarchies had 21different words, and the target conceptualhierarchies had from 25 to 30 different words,averaging 27. The NR measure thus gives arough index of associative variety relative tothe target hierarchies. The final column,labeled Recall, gives the average recallabilityof these hierarchies when theywere presentedin blocked form, pooling all four trials fromExp. IV for the associative hierarchies and the

TABLE 7

Mean Proportion Correct(p), AgreementIndex (Al), Total Number

of

Different Responses (NR), and Recall

for

the 12 Hierarchies

Al=—, r=T7

Level 2 Level 3 Level 4 Total

Associativehierarchy jp Al p Al p Al NR Recall

SaltCheeseHammerWish

.33

.10

.43

.37

.74

.44

.89

.81

.03

.20.20.20

.61

.78

.78

.80

.27

.14,09.14

.83

.58

.64

.72

.21

.18

.18.21

115134126123

.74.77.79.81

Average .31 .72 .16 .74 .16 .69 .19 124 .78

Conceptualhierarchy P AI P Al P Al P NR Recall

InstrumentsMineralsBodyPlantAnimalTransportationOccupationClothing

.300

.23

.060.70.20.20

.67

.56

.48

.50

.48.96.63.39

.23.02.55.07.22.13.16.23

.76

.64

.96

.63

.72

.81

.78

.85

.42.51.48.2')

.27

.21

.34

.36

.93

.95

.94

.76

.90

.86

.90

.96

.37.39.47.24.24.27.30.35

14512010316014711714095

.71.74.81.77.72.72.64.71

Average .21 .58 .20 .77 .36 .90 .33 128 .73

339HIERARCHICAL RETRIEVAL SCHEMES

first two trials pooling Exps. I and II for theconceptual hierarchies.

The scores in Table 7 vary widely and fallinto few discernible patterns. For the con-ceptualhierarchies, theproportion correct andthe agreement indices increase at deeperlevelsof the tree. For the associative hierarchies,the average proportion correct decreasessomewhat at deeper levels of the tree, but theaverage agreement index remains constantacross levels. These patterns ofchanges in Alfor the conceptual vs. associative hierarchiesare understandable in terms of the increasingconstraints upon responses imposed by thegreater contextual information at the deeperlevelsof theconceptual but not the associativehierarchies. The overall proportion correct ishigher for the conceptual hierarchies, but thisis due to the preponderance of level-4 wordswhich are more constrained and better guessedin the conceptual hierarchies.

The NR or associative variety measurereveals the open-ended or nonexhaustivecharacter of all these hierarchies. The numberof different responses given by ten Ss isconsistently 4-7 times more than the numberof words in the hierarchies, with the propor-tions in excess of the target words beingsomewhat higher for the associative than forthe conceptual hierarchies. Again, this seemsattributable in part to the greater constraintsin the conceptual hierarchies imposed by therestricted inclusion or part-whole relationbetween successive nodes.

Finally, we examine the relation betweenrecallability of the hierarchies and someoverall index ofnaive Ss' ability to generate thedifferent hierarchies. The meanrecall probabi-lities for the blocked hierarchies areshown inthefinal column ofTable 7. For the associativehierarchies, there is only very small variationin recallability and in overall generationprobabilityfor the four hierarchies. Becausethe range exhibited in both variables is smallrelative to theprobable error variance of thesemeasures,no systematic correlation of the twovariables could possibly be shown in this

situation. A similar statement applies forrelating generation probability of the associa-tive hierarchies to their recallability from arandom list; the range of variation in recallfrom the randomly presented lists (Exp. IV)was only 5%, which is too small to considercorrelating that variable with generationprobability.

Turning to the conceptual hierarchies, thesituation there is somewhat better for investi-gating the relation between generationproba-bility and recall, since there was largervariation in both factors. However, thecovariation in generation probability andrecallability is not at all strong in this case.The body hierarchy was highest on bothvariables, but there were several large dis-crepancies in rankings of other hierarchies onthe two variables. If the eight hierarchies arerank ordered onboth variables, thecorrelationof therank orders is only+. 1 4,indicating a veryweak relationship between the two variables.The Pearson correlation coefficient betweenthe two variables is +.40, but this is well belowthe value of .62 required to reject the nullhypothesis of zerocorrelation at the 5% level.The generationprobabilities of theconceptualhierarchies were also correlated with therecallability of these words when they werepresented in random order (condition WRin Exp. I and II). Here, too, the correlationswere low and insignificant: Rank-orderp = +.24; Pearson r = +A0 but seven of theeight hierarchies differed by only 3% inaveragerecallability. These results show thatrecallability of the individual hierarchies isnot predictable from the probabilities withwhich naive Ss can generate the target hier-archies. In a similar vein, recallability of theconceptual hierarchies correlates very poorly(p = +.15) with the NR index of associativevariety. Possibly other indices of hierarchicintegration or association could be composedwhich will correlate better with recallability ofthese 12 hierarchies; but the clear failings ofthese obvious variables prognosticate littlesuccess for such a search.

340

BOWER

ET AL.

One might entertain the hypothesis that theobserved differences in free recall betweenthe Blocked vs. Random input conditions intheprior experiments simply reflects the differ-ential guessing of the words when S does ordoes not know the principle of hierarchicconstruction. For example, in Exp. IV with theassociative hierarchies, the Blocked vs.Random difference in recall, averagedover thefour trials, was .19, which is exactly theaverage generation probability for thesehierarchies in Table 7. However, there areseveral deficiencies in this account of theBlocked vs. Random differences in recall.First, with the conceptual hierarchies, theaverage generation probability (.33) con-siderably underestimates the recall advantageof the Blocked list on Trials 1 and 2 (averagedifference ca. .48). Second, it is not true thatRandom Ss would never guess words on thesehierarchies, since the general categoricalnature of the word sets was still fairly obviousto Ss receiving the random presentations, asshown by the categorical nature of the intru-sions by these Ss; so this factor would tend toreduce the guessing differential below the .19or .33 figures of Table 7. Third, if Ss werereally guessing words from the conceptualhierarchies, then one should find a very largenumber ofcategorical intrusions, since the NRmeasure in Table 7shows that these hierarchiesare very open-ended. In fact, however, recallintrusions never exceeded 1 % in either theBlocked or Random condition on any trial.

The upshot of these considerations is thatthe differential recall produced by the Blockedvs. Random presentations is not consistentwith a differential guessing probability exhi-bited by Ss who do vs. Ss who do notknow theprinciple ofhierarchic construction. Nor is therecallabilityof a given hierarchy, with eitherBlocked or Random presentation, related inany significant degree to the average probabi-lity with which a naive S can generate thewords in that hierarchy from theconstructionprinciple. We are therefore left with thetheoretical reconstruction given earlier for

explaining the effect of Blocked vs. Randompresentation; namely, the Blocked spatialarray enables S to directly strengthen parti-cular category-instance associations and italso provides him with a systematic (hier-archic) plan for cuing these candidates forrecall, which candidates are recalled only ifthey pass a recognition test for list member-ship.

Discussion

The message of these studies is simple: If Scan discover or learn a simple rule orprinciplewhich characterizes the items on a list andwhich relates them to one another, then heuses thatrule as aretrieval plan inreconstruct-ing the items from memory,with a consequentimprovement in his performance. The prin-ciple characterizing a hierarchy is that ofrecursive rewriting by associative transitions.For ourconceptual hierarchies, the associativetransitions are primarily of one restrictedtype,namely, therelationship of class inclusionor part-whole. But Exp. IV showed facilitationeven when the base-type of associative transi-tion varied widely over the set of words. Thepotent effect of such retrieval plans on freerecall was illustrated in Exp. I, where Ss wererecalling 1 12 words perfectly after twoor threeinput trials, which may be compared with anynumber of other reports in the literatureshowing much poorer recall of shorter, lessorganizedlists.

A question to which we must return iswhether this result is trivial, whether it isanalogous in some obscure way to asking Sto "recall" the numbers from 1 to 112 Toanswer this objection adequately wouldrequire a fairly detailed analysis of the opera-tions involved in "memoryexperiments" (freerecall in particular), the involvement ofresponse-generation rules in suchexperiments,and the difference between "remembering"items from a presented set versus generatingthose items given only knowledge of a rulecharacterizing the set. These are difficultquestions, but we would nominate them as

HIERARCHICAL RETRIEVAL SCHEMES 341

extremely important questions to answer ifonewants to know what is involved in the"remembering" seen in free recall.

The observable behavioral difference be-tween remembering a set of items andgenerating a list according to aprinciple is inthe nature and extent of intrusion errors. Ofcourse, whether a particular candidate gene-rated by the rule is or is not an intrusiondepends on how well the rule generates onlythose items on the list and no others. To theextent that the rule fails to discriminate wellbetween list and nonlist candidates, to thatextent specific item-memory has to be usedalong with the rule to make it an adequaterecall device. And it has beenconjectured herethat this item-memory is essentially "occur-rence" recognition, mediatedby some kind ofrecency or contextual frequency tag storedalong with a lexical item in S's long-termsemantic memory. The evidence collected inExp. V tends to discredit the view thatBlockedSsare "merely guessing"from their long-termsemantic memory in a mannerconsistent withtheir knowledge of the list-constructionprinciple.

It is of interest to compare the presentexperiments to those of Whitman and Garner(1962) and Miller (1958) which also involvedfree recall of rule-generated lists. In Miller'sexperiment, Ss saw many strings of 3-5 lettersand then tried to recall them. For one list,the strings were a subset of those produced bya simple Markov generator,which is a set ofleft-to-right rewriting rules for a finitevocabulary. For example, the three rewriterules A -* B, B -> (C orB), and C -> (A or C),will generate strings like AB, BBC, CCAB,BCCABBC but not CB, BA, or BAC. Millerfound that a list of such rule-generated stringswas more easily learned and recalled than wasa list of random strings of comparable lengthwith the same letters. If the list subset had hada simple structure (e.g., all permissible stringsof length 4 beginning with C), thenknowledgeof the rewrite rules would have permitted Sto "recall" all the presented strings. But in

Miller's experiment, the list subset was notespecially restricted (different initial lettersand strings of differing lengths) and S did notknow the principle of list construction (therewrite rules) at the start of the task. However,there can be little doubt that the recallimprovement observed with practice was inpart due to the S learning some or all of therewrite rules exemplified by the list strings(see also Smith, 1966).

The experiments by Whitman and Garner(1962) (also Garner and Whitman, 1965) arerelevantbecause their Ss in principle knew howto generateall possible items of thepopulationfrom which a subset was to be recalled. Atypical task might expose S to eight of thepossible 16 geometric figures that could becomposed by combiningfour binary attributes(e.g., one or two, large or small, red or blue,circles or squares). The principle for con-structing thepopulation ofpatterns was simpleenough that most Ss could probably generateall 16 patterns in the population after only afew exposures. The problem for thefree-recallS in such experiments is not oneinvolving aninability to generateitems on thepresented list,but rather one of suppressing intrusions, ofinferring a restricting rule which will generateonly those patterns presented and no othersof the population. Whitman and Garnershowed that the difficulty of free recall in thissituation dependedupon thecomplexityof therestrictive rule characterizing the subset ofpresented patterns. For instance, if thepresented subset were to occupy an elementarypartition of the population (e.g., the eight redfigures), then free recall would be extremelyeasy. Whitman and Garner also point out thebasic similarity between such recall experi-ments andconcept learningexperiments, ifoneidentifies the presented subset for free recallwith the list of "positive instances" of aconcept experiment. A rule for generatingfree recall of the presented list would have asan essential ingredient the concept charac-terizing the members of that subset of thetotal population.

342

BOWER

ET AL.

The relevance of these studiesby Miller andWhitman and Garner to our ownexperimentsis relatively direct. In each experiment, thepresented items were characterized by somestructural principle which could be used togenerate these items, and learning was largelya matter of discovering the rules and thenrecognizing the presented from nonpresentedcandidates that could be generatedby therulesapplied to the base vocabulary. In eachexperiment, of course, the difficultyof recalldepended upon the structural characteristicsof the whole list and was not some simplecombination of the difficulty of the individualitems considered in isolation. One might saythat the availabilityor recallability of a givenitem depended upon "wholist" properties ofthe list in which it was embedded.

According to the viewpoint espoused here,free recall is mediated by S using a retrievalplan for cuing or generating plausible candi-dates torecall, and an "executiveeditor" whichchecks these candidates for recency recogni-tion before overtly recalling them. Dale ( 1967),Kintsch (1968) and others have proposedsimilar views. The last few input items in theecho box of short-term memory may berecalled directly without mediated cuing, butthe hypothesis supposes that most of theother items in free recall would be generatedby a cuing system. The cuing system could bea set of rules, some structural informationabout thecomposition of the list,an alphabeticscheme (Earhard, 1967), or a pegword system(Wood, 1967a). Last but not least theeffectivecue for recall of a given word may be priorrecall of other words. But even this has toregress eventually back to an implicit cuingsystem. The improvement in free recall overmultiple trials wouldresult, on this view,fromseveral factors: (a) more discriminatingoccurrence-information being stored for listwords, (b) increasing integration of subjectiveclusters or groups of words which S treats asunits, and (c) developmentof a more adequateretrieval plan for S to cue recall of his subjec-tive units. The characteristic of thecuing plan

is that it should be afamiliar oreasily learned,summary abbreviation suggesting or leadingto the words on the list. The word "plan"need not suggest any elaborate mentalistic orcognitive reconstruction of behavior: Millen-sen (1967) and Suppes (1968) have shown howhierarchical TOTE units (cf. Miller, Galanter,and Pribram, 1960) can be analyzed in termsof conditional implicit stimulus-responseconnections, so this language is noncommittalwith respect to S-R analyses ofrecall.

The evidence to date, collected when S isgiven a systematicretrieval plan or mnemonic,tells us that such plans aresufficient to producevery high levels of recall. The evidence doesnot yet prove that such a plan is necessary orrequired for S to produce high levels ofrecall,but this hypothesis is beginning to lookincreasingly attractive.

References

Anisfield,

M., and Knapp, M. Association,synony-mity, and directionality in false recognition.Journal of Experimental Psychology, 1968, 77,171-179.

Bower, G. H. A multicomponent theory of thememory trace. In K. W. Spence and J. T. Spence(Eds.) The psychology of learning and motivation.Vol. 1. New York: Academic Press, 1967.

Cofer,

C. N., Bruce, D. R., and Reicher, G. M.Clustering in free recall as a function of certainmethodological variables.Journal ofExperimentalPsychology, 1966,71, 858-866.

Cohen, B. H. Some-or-none characteristicsofcoding.Journal of Verbal Learning & Verbal Behavior,1966,5, 182-187.

Cohen,

B. H., and

Bousfield,

W. A. The effects of adual-level stimulus-word list on the occurrence ofclustering in recall. Journalof GeneralPsychology,1956,55, 51-58.

Dale, H. C. A. Familiarity and free recall. QuarterlyJournal of Experimental Psychology, 1967, 19,103-108.

Earhard, M. Cued recall andfree recall as a functionof the numberof items per cue. Journal ofVerbalLearning & VerbalBehavior, 1967, 6, 257-263.

Garner, W. T., and Whitman,J. R. Form and amountof internal structure as factors in free-recalllearning of nonsense words. Journal of VerbalLearning & VerbalBehavior, 1965, 4, 257-266.

343HIERARCHICAL RETRIEVAL

SCHEMES

I

Kintsch, W. Recognition and free recall of organizedlists. Journal of Experimental Psychology, 1968,

78,481-487.

Mandler, G. Organization and memory. In K. W.Spence and J. T. Spence (Eds.) The psychology oflearning and motivation, Vol. 1, New York:AcademicPress, 1967.

Mandler, G. Association and organization: Facts,fancies, and theories. In T. R. Dixon and D. L.Horton (Eds.) Verbal behavior and generalbehavior theory. Englewood

Cliffs,

N.J.: Prentice-Hall, Inc., 1968.

Millenson, J. R. An isomorphism between stimulus-response notation and information-processingflow diagrams. Psychological Record, 1967, 17,305-319.

Miller, G. A. The free recall of redundant strings ofletters. Journal ofExperimentalPsychology, 1958,56,485-491.

Miller, G. A.,

Galanter,

E., and Pribram, K. Plansand the structure of behavior. New York : Holt,Rinehart & Winston, Inc., 1960.

Miller, G. A., and Selfridge, J. Verbal context andthe recall of meaningful material. AmericanJournal ofPsychology, 1950,63, 176-185.

Paivio, A., Yuille, J.

C,

and Madigan, S. A.

Concreteness,

imagery,and meaningfulness valuesfor 925 nouns. JournalofExperimentalPsychologyMonograph Supplement, 1968, 76, No. 1, Part 2,1-25.

Parks, T. E. Signal detectability theory ofrecognitionmemory performance. Psychological Review,1966,73,44-58.

Postman, L. The present status of interference theory.In C. N. Cofer (Ed.) Verbal learning and verbalbehavior. New York:

McGraw-Hill,

1961.

Postman, L., and Keppel, G. Retroactive inhibitionin free recall. JournalofExperimentalPsychology,1967,74,203-211.

Smith, K. H. Grammatical intrusions in the free recallofstructuralletterpairs. JournalofVerbalLearning& VerbalBehavior, 1966, 5, 447-454.

Suppes,P. Stimulus-responsetheory offinite automata.Technical Report 133,Psychology

Series,

Institutefor Mathematical Studies in the Social

Sciences,

Stanford University, Stanford, Cal., 1968.Tulving, E., and Pearlstone, Z. Availability versus

accessibility ofinformation in memory for words.Journal of Verbal Learning & Verbal Behavior,1966,5,381-391.

Tulving, E., and Thornton, G. B. Interaction be-tween proaction and retroaction in short-termretention. Canadian Journal of Psychology, 1959,13, 255-265.

Underwood, B. J. False recognition by implicitverbal responses. JournalofExperimentalPsycho-logy, 1965,70, 122-129.

Whitman, J. R., and

Garner,

W. R. Free recalllearning of visual figures as a function of form ofinternal structure. Journal of ExperimentalPsychology, 1962, 64, 558-564.

Wood, G. Mnemonic systems in recall. Journal ofEducational Psychology Monograph Supplement,1967a,58, no. 6, Part 2, 1-27.

Wood, G. Category names as cues for the recall ofcategory instances. Psychonomic

Science,

1967b,9, 323-324.

Wood,

G.,

and Underwood, B. J. Implicit responsesand conceptual similarity. Journal of VerbalLearning & VerbalBehavior, 1967,6, 1-10.

(Received November 8, 1968)