See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223321611

Hierarchical Retrieval Schemes in Recall of Categorized Word Lists

Article in Journal of Verbal Learning and Verbal Behavior · June 1969

DOI: 10.1016/S0022-5371(69)80124-6

CITATIONS READS
505 13,348

4 authors, including:

Gordon Bower Alan M. Lesgold

Stanford University University of Pittsburgh
215 PUBLICATIONS 31,743 CITATIONS 95 PUBLICATIONS 4,246 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Gordon Bower on 18 April 2018.

The user has requested enhancement of the downloaded file.

JOURNAL OF VERBALLEARNINGAND VERBALBEHAVIOR8, 323-343 (1969)

Hierarchical Retrieval Schemes in Recall

of Categorized Word Lists1

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 upon

their free recall. Ss 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 similar with associative as well as conceptual
hierarchies, (b) was attenuated with recognition tests of memory, and (c) could not be
accounted for by associative "guessing." Another experiment demonstrated retroactive
facilitation in recall of List 1 when List 2 contained the hierarchic superordinates of the
words on List 1. Analyses suggest that the hierarchic principle was used as a retrieval plan
for cuing recall, with generated candidates monitored for their list membership before being
overtly recalled.

The following experiments were undertaken order, whereas in random presentation the
to demonstrate the influence of structural several instances of each category are sepa-
organization upon the free recall of conceptual rated by many intervening items in the
word hierarchies. The studies were initiated exposure list. Psychologically, blocking of
in the belief that previous free recall experi- instances at input is more likely to lead S to
ments have investigated only relatively weak discover the superordinate category (cf. Wood
manipulations of this structural variable, and and Underwood, 1967) and thus it should
the further belief that it should be possible to mediate better recall. Although blocked lists
arrange for a much more potent demonstra- are often recalled better than random lists,
tion. the effect has often been surprisingly small.
This research is most relevant to previous For example, in the Cofer et al. study, blocking
studies of free recall of categorized word lists. produced an average recall advantage of only
If the list consists of several instances of 13 % over a randomly ordered list. In Cohen's
several taxonomic categories (e.g., animals, more recent and extensive studies, with long
colors, occupations, etc.), it will be better lists of many categories, there has been no
recalled than a comparable list of unrelated effect at all of blocking vs. random input upon
words; moreover, in his recall, S tends to total recall.
group together words from the same category. If one believes that structural information
A number of variables inherent in this situation about the input list is an important component
have been studied (e.g., Cofer, Bruce, and of S's ability to recall it, then the small or
Reicher, 1966; Cohen, 1966). One of imme- non-existent effect of blocking in these prior
diate relevance to our studies is the presen- studies is rather disheartening. However,
tation-order of a categorized word list. In the possibility remains that the structural-
blocked presentation, the instances of a given organization variable has been only weakly and
category are presented in adjacent temporal ineffectively manipulated in the prior studies.
In consequence, one searches for a more
i This research was supported by a grant (M-13950) potent way to manipulate this variable, and
to the first author from the National Institute of this is what we have done in the following
Mental Health. experiments.
11 323
© 1969by AcademicPress Inc.
324 BOWER E T AL.

LEVEL

i I ~I E R A

Platinum Aluminum Bronze Sapphire Limestone

Silver Copper Steel Emerald Granite

Gold Lead Brass Diamond Marble
Iron Ruby ~ate

FIG. 1. The "minerals" conceptual hierarchy.

We have made three obvious changes from language, and in actuality there are a very
the prior studies of category blocking. First, large number of plausible options or branching
the list to be recalled contains not only points in generating such a hierarchical tree.
instances of a category, but also the category Because of the open, nonexhaustive character
labelitself. If recall of the category label serves of such conceptual word hierarchies, the one
as a cue for recall of its instances (cf. Tulving actually used can be considered as only one
and Pearlstone, 1966), then the presence of of many plausible trees that could have been
the category name on the recall list should developed from the level-1 word. In formal
boost the recall of the structured material. terms, the trees used in these experiments vary
Second, the word list is selected to be a considerably from one another in the number
hierarchically organized set of nested cate- of nodes at each level, and the use of nouns
gories, so that the "instances" of a high-level versus class-restricting adjectives at the higher
category serve in turn as superordinates for levels. Many of the words have ambigu-
still lower-level instances. We have constructed ous meanings, so that the intended sense
about eight of these conceptual hierarchies, of the word would be established only
one of which (for minerals) is shown in Fig. 1. by seeing it in relation to its hierarchical
The four different "levels" of the hierarchical context.
tree are indexed by the left-hand column of The third change we have made from the
numbers. prior studies is in the method of presenta-
Such word trees have an obvious structure tion. The prior studies have typically
defined generally by the "class inclusion" presented the lists one word at a time, whereas
relationship, and there would probably be we present a complete set of Words all at once
fairly general agreement on the appropriate- for prolonged exposure. The one-at-a-time
ness of the words at different levels of the tree. method has the advantage of operationally
Surprisingly, however, we have found that equating exposure time to each word, whereas
naive Ss have considerable difficulty in trying the complete-presentation method can only
to generate such trees given only the level-1 equate total time for the entire set of words.
word and a general characterization of the However, the method of complete presentation
target hierarchies to be generated (cf. our makes it easier for S to discover more struc-
Exp. V later). The notion of "levels of class tural information about the input list than
inclusion" turns out to be a terribly vague and does the one-at-a-time method, and the former
imprecise concept when applied to natural has been used for this reason.
 HIERARCHICALRETRIEVALSCHEMES 325

EXPERIMENT I was used even for the Random lists. After seeing four
such cards, S recalled the words orally in any order he
O u r first e x p e r i m e n t is in essentials a simple
preferred. Total time allowed for recall was five seconds
c o m p a r i s o n o f free recall o f hierarchical w o r d per word on the input list. No S ever needed this much
lists t h a t are presented in a b l o c k e d as o p p o s e d time, and S typically initiated the next input trial by
to a r a n d o m i z e d fashion. F o u r w o r d hier- indicating that he could recall no more.
archies were l ear n e d concurrently. T h e Ss in The six experimental conditions were as follows:
(a) Whole Blocked (WB)." The four complete hier-
the B l o ck ed c o n d i t i o n were exposed to the
archies (averaging 28 words p e r hierarchy) were
f o u r c o n c e p t u a l hierarchies o r g a n iz e d in presented in conceptually blocked fashion, one per
vertical trees as s h o w n in Fig. 1. F o r Ss in the card, for all four trials. (b) Whole Random (WR):
R a n d o m condition, the same w o r d s were The same set of 112 words, approximately 28 on each
t h o r o u g h l y scrambled, then assigned ran- of four cards, were presented on each trial. The nodes
in the tree on each Card were filled by random selection
d o m l y to the nodes o f four spatial trees. T h e
without replacement from the 112 words, avoiding
spatial tree seen by R a n d o m Ss h a d no obvious conceptual relations amongst words in
a p p a r e n t c o n c e p t u a l significance since the successive nodes of a tree. (c) Progressive Blocked (PB):
Words l o cat ed a b o v e a n d b e lo w a given node- On Trial 1, each exposure card contained only the
w o r d bore no o b v i o u s c o n c e p t u a l relation to levels 1 and 2 words (14 in total) in conceptually blocked
fashion. On Trial 2, the appropriate level-3 words were
the node.
added below the levels 1 and 2 words on each card
A subsidiary factor that was varied ortho- (40 in total); and on Trials 3 and 4, the 72 level-4
gonal to the Blocked versus R a n d o m f a c t o r words were added appropriately. Thus on Trials 3 and
was Progressive parts versus W h o l e presen- 4, the exposure conditions for PB Ss were identical
t a t i o n o f the hierarchies. In the Progressive to those of WB Ss. (d) Progressive Random 1 (PRI):
This used the same trees as the WR condition, but
parts condition, the level-1 a n d -2 words were
with progressive exposure of levels 1 and 2 words on
presented on Trial 1, levels 1, 2 and 3 on Trial 1, levels 1, 2, and 3 on Trial 2, then the full
Trial 2, then the full set (levels 1-4) on Trials 4-level hierarchies on Trials 3 and 4. (e) Progressive
3 an d 4. T h e t h o u g h t here was that recall o f Random 2 (PR2): This was similar to PR1 except the
the new level-4 w o r d s on Trial 3 m i g h t be randomization of words was within, not across, levels
of the conceptual trees. A level-n word in a conceptual
a p p r eci ab l y aided by S h a v i n g already h a d
tree was used only at a level-n node in a PR2 tree; the
practice in recalling the B lo c k e d level 1-3 difference between such random trees and the concep-
s u p e r o r d i n a t e words. N o such benefit w o u l d tual trees is only in the vertical relationships among
be expected for the Progressive parts Ss words above and below the nodal words. Thus the
receiving the R a n d o m list. PR2 word sets being recalled over trials were identical
to those of condition PB, except that the scrambling
of words across cards did not allow the PR2 Ss to
Method
easily disc0;~er the conceptual relationships. (jr) Pro-
Design and procedure. There were six groups of Ss, gressive Unrelated (PU), These Ss learned 112 un-
four learning by Progressive parts and two by the related nouns by the progressive parts method, with
Whole presentation methods. Each S had four input- 14 on Trial 1, 40 on Trial 2, and all 112 on Trials 3
output trials on a list composed eventually of four and 4. The words were arrayed in vertical tree form
complete conceptual hierarchies comprising 112 words. as in the other cases. This condition was run for
Because of the much shorter times required for Trials 1 comparison to the PR conditions which had conceptual
and 2 in the Progressive conditions, we were able to word lists. The 112 unrelated nouns were comparable
run a complete replication of the experiment with in Thorndike-Lorge frequency to the words of the
these Ss within the experimental hour. The replication conceptual list, but because they were all nouns, tended
simply maintained the same progressive-parts condi- to have higher "concreteness" ratings (Paivio, Yuille
tions but with a different set of four hierarchies. and Madigan, 1968) than did the words on the
The Ss were run individually. A complete or partial conceptual list.
hierarchy was shown printed on a large 5 x 8-in. card As indicated before, there was time to complete
with a total study time calculated at two seconds per four trials on a second list of four different hierarchies
word on the card. The words on each card were arrayed with Ss in the Progressive conditions. The two sets of
in the form of a vertical tree but without the connecting four hierarchies were animals, clothing, transportation,
lines and circles shown in Fig. 1. The same tree-form and occupations in set 1, and plants, instruments
326 BOWER ET AL.

body parts, and minerals in set 2. Half of the Ss in the Blocked condition is seen again to be m u c h
each progressive condition had set 1 first and set 2 higher than in the comparable R a n d o m
second, while the remaining Ss had the reverse order.
Similarly, four of the eight Ss in the WB and WR conditions. Recall scores in the two R a n d o m
conditions learned set 1 and four learned set 2. conditions, PR1 and PR2, do not differ
The Ss were 48 Stanford undergraduates fulfilling significantly f r o m each other, nor do they
a service requirement for their introductory psychology differ significantly f r o m recall of Ss learning
course. There were eight Ss in each of the six experi- the unrelated word lists. All three o f these
mental conditions.
conditions are significantly inferior to condi-
Results tion PB f r o m Trial 2 onwards.
Performance on Trials 3 and 4 provides
F o r the four Progressive groups, there was
comparisons between the Whole versus
no significant different in performance com-
Progressive parts methods, since on these
paring their List 1 with List 2, so their
trials Ss were exposed to the same full list.
performance on b o t h lists was pooled to
Within the Blocked conditions, W B signifi-
increase reliability o f the following analyses.
cantly exceeded PB on Trial 3 simply because
there was no variance' for W B Ss on this trial
TABLE 1
(all recalled perfectly). Within the R a n d o m
AVERAGE WORDS RECALLEDOVER FOUR TRIALS
conditions, recall in W R exceeds that o f
Trials either PR1 or P R 2 on both Trials 3 and 4.
Thus, in terms o f overall recall in these condi-
Condition 1 2 3 4 tions, it was clearly advantageous to present
Whole: all words on all trials rather than progressively
Words presented 112 112 112 112 increasing the exposure set over trials.
Blocked 73.0 106.1 112 112 One m a y ask the further question whether
Random 20.6 38.9 5 2 . 8 7 0 . 1 prior practice at recalling the level 1-3 words
Progressive: substantially improves S's ability to recall the
Words presented 14 38.5 112 112 level-4 words when they are presented for the
Blocked 13.0 34.5 93.2 108.4 first time on Trials 3 and 4. The level-4 recall
Random 1 8.6 16.0 "34.9 46.3 data are presented for the relevant groups in
Random 2 8.4 15.5 33.9 52.6
Unrelated 9.9 18.7 32.4 51.1 Table 2.
Level-4 recall for g r o u p PB on Trials 3 and
4 significantly exceeds level-4 recall for group
The average recall scores for the four trials
for the six conditions are shown in Table 1. TABLE 2
L o o k i n g first at the two Whole-presentation RECALLOF THELEVEL-FouRWORDS
groups, recall in condition W B is seen to be
Trials
markedly superior to that o f g r o u p W R . M e a n
recall is 3.5 times better in condition W B on Condition 1 2 3 4
Trial 1, and there is no overlap a m o n g the Whole:
recall scores o f the two groups o f Ss on any Words presented 73.5 73.5 73.5 73.5
trial. Recall of the 112 words in condition W B Blocked 44.6 69.6 73.5 73.5
is almost perfect by Trial 2. Obviously, the Random 13.2 22.8 35.2 44.8
structural organization o f the blocked input Progressive:
list had a tremendously powerful effect on free Words presented 0 0 73.5 73.5
recall in this situation. Blocked -- -- 56.6 70.4
Turning next to the four Progressive condi- Random 2 -- -- 16.3 30.0
Unrelated -- -- 13.8 27.6
tions in the lower portion of Table 1, recall in
 HIERARCHICAL RETRIEVAL SCHEMES 327

WB on Trials 1 and 2 (t = 2.17, p < .05), so and the facts above, this is a credible
that proactive facilitation exists when the new cl/tim.
words to be recalled fit as subordinates into A question related to the results above is
the conceptual structure practiced on Trials 1 whether recall of a level-n word for WB Ss
and 2. However, a similar facilitation appears serves a cuing function, being correlated with
comparing Trials 3 and 4 of group PR2 with recall of the level-n+l instances nested within
Trials 1 and 2 of group WR; a t test for the it. To answer this question, we considered the
level-4 recall over the respective trials for the level-4 recall conditional upon level-3 recall
two groups is significant, t(14) = 2.63, p < .05. by WB Ss on Trial 1 Where there were the most
For the unrelated words, comparing Trials 3 data and recall failures. The average prob-
and 4 of group PU with Trials 1 and 2 of ability of recalling a level-4 word was .66 when
group WR, the difference in recall is in the its level-3 word was recalled, but only .30 when
same direction but not significant. From these its level-3 word was not recalled. Moreover,
several comparisons, the overall conclusion is when a level-3 word and at least one of its
that recall of a new subset of words presented level-4 words were recalled, in 90 ~ of these
for the first time within a larger set is facilitated cases recall of the level-3 word preceded recall
if there has been prior practice at recalling of the first level-4 word (not necessarily in
the other words. This result may be specific immediate succession, however). Those cases
only to the complete presentation method used where level-4 words were recalled without
here, since that permits S to spend relatively prior recall of the level-3 word were most often
more time studying the new words added to instances in which the level-4 words were cued
his input list. (preceded) by a level-2 word, skipping the
Further analyses of recall protocols demon- level-3 word. These results accord with the
strated that WB Ss were essentially using the expectation that recall of a word served to cue
structural information in the blocked input recall of the corresponding words at lower
as a retrieval plan for generating their recall. levels. When a nodal word was not recalled,
First, in condition WB, Ss' recall was organ- the entire "tree" developing out of that node
ized almost entirely according to conceptual was likely to be missing in recall.
categories, whereas Ss in condition W R tended In contrast to this strong dependence
to show conflict between clustering according between recalls of conceptual levels 3 and 4
to the underlying conceptual categories versus for the blocked hierarchies, recall of these same
according to the spatially contiguous words words was virtually independent for Ss who
in the randomized input. Second, all WB S s viewed random hierarchies. Scoring Trial 1
recalled their hierarchies from level-1 down recall of W R Ss for the same words as above,
through level-4 words (from top down), and the conditional probability of recalling a
never in the reverse order. In doing this, some level-4 word was .23 when its level-3 word was
Ss in condition WB generated the conceptual recalled, and was. 17 when its level-3 word was
trees in breadth across levels (i.e., all the not recalled. These probabilities reveal no
level-n words of a hierarchy before any of its contingencies in excess of chance expectation
level-n+l words), whereas other Ss generated on the null hypothesis of independent events,
the hierarchy in depth, along a given (say) x2(1) = 1.5, p > .30. Thus, the influence on
left-going branch before returning to higher recall of the moderately strong normative
nodes to generate in depth the other branches associations from a level-n category word to a
from it. Most WB and PB Ss claimed that they level-n+l instance was considerably modu-
could have reproduced the hierarchical arrays lated by the method of presentation. If the
exactly a~ shown had they been permitted to associated words were presented in close
write their recall; given their perfect oral recall temporal and spatial contiguity, so that S
328 BOWER E T AL.

might rehease them together, they tended to 40-word list classifiable into eight independent
be recalled together; if they appeared widely categories of five instances each, versus
separated in a random list, then the same (b) a dual-level list consisting of four super-
words were recalled almost independently in ordinate categories each divided into two
this case. subordinate categories (e.g. ,feline animals and
There were some small differences in the canine animals), with five instances in each
recallability of the eight conceptual hier- of these eight subordinate categories. Cohen
archies. Only the WB and WR conditions on and Bousfield reported that total recall of
Trials 1 and 2 were considered for these these two lists was very similar (17.6 vs. 18.1
comparisons since, by the time the full hier- out of 40), but that recall clustering differed
archies were presented, the PB Ss were making somewhat. The Cohen and Bousfield study
very few errors. The proportions of words differed in a number of respects from ours.
recalled for the various hierarchies averaged For example, in their study the category names
over Trials 1 and 2 for WB (and WR) Ss were were neither presented nor recalled, and items
as follows: Plants, 87(26); minerals, 85(21); were presented singly in random order. These
body parts, 84(34); transportation, 81(16); procedures surely would reduce the probabi-
instruments, 79(27); occupations, 77(25); lity that S would notice and utilize the dual-
animals, 76(24); clothing, 72(29). The range level organization of the second list. In this
of variation in recall of the word sets is about regard, we may point out that our PR Ss,
15 ~o for bothgroups, but the rank ordering of who received a hierarchical word list but
the sets by the two groups is quite discrepant randomized, did not recall any better than our
(rank correlation = - . 1 4 ) . The poor correla- P U S s who had unrelated words. Obviously,
tion of the recall ranking of the eight hier- the list-structure has to be discovered and
archies in conditions WB and WR may have utilized if S is to derive any benefit from it in
been due to treatment effects or t o low his recall.
reliability since only four Ss in a given Complete presentation of the Blocked
condition learned a given hierarchy. hierarchies provide s S with a lot of structural
Perhaps the main fact to be remembered information about the word list--he does not
from Exp. 1 is that complete presentation of have to discover it for himself. This structural
these conceptual hierarchies produced a information in turn provides S with a plan
tremendous facilitation of free recall relative for retrieving the words from memory. The
to a random ordering of the same words. The salient characteristics of a retrieval plan are
effect is similar to what one finds in free recall that it tells S where to begin his recall, how to
of a sentence versus scrambled words (e.g., proceed systematically from one unit to the
Miller and Selfridge, 1950). In both cases, next and to the words within each unit. The
recall depends upon the way the words are plan also helps S to monitor the adequacy of
arranged, with a familiar structure in one case his recall, helping him to identify where parts
contrasted to unfathomable randomness in are missing and to identify when he has
the other. finished. It is plausible that a central ingredient
The notion of hierarchical organization as a in the present hierarchical retrieval plan is
recall aid is hardly original with us. Mandler associative cuing of the words at level-n+l by
(1967, 1968) has hypothesized that recall is recall of the superordinate category at level-n.
largely a matter of subsuming list items under Having already recalled masonry (stones), S is
a hierarchical array of categories. Some set to search for recency-tags on words in his
earlier experimental work was done by Cohen associative hierarchy to this category (e.g.,
and Bousfield (1956) who were interested in granite, limestone, sandstone, flagstone, etc.).
recall clustering produced by either (a) a It is further plausible to assume that the
 HIERARCHICAL RETRIEVAL SCHEMES 329

implicit candidates generated to the category processes. In terms of the theory outlined
cue are recalled only if they pass a recognition above, the occurrence of an item on the input
criterion as having been on the list. Thus, of list causes information of some kind (a tag
the four masonry-stone candidates, S might denoting recency, frequency, or trace strength)
overtly recall granite and limestone but not to be attached to the representation of that
intrude sandstone and flagstone since the latter word in S's semantic memory. In a test f o r
two words were not tagged as having been seen recognition (or discrimination of list member-
recently. ship), the occurrence-information attached to
The detailed analyses of the recall protocols the test-word in memory is consulted for a
support this general view. When recalled, decision, yielding a judgment of recognition
the superordinate oategory almost always only if the information there exceeds a criterion
preceded recall of its instances, and instance (e.g., Bower, 1967; Parks, 1966). Such
recall was considerably poorer if the super- recognition tests, which directly provide the
ordinate was not recalled. Moreover, the test word, clearly bypass the search and
intrusion data in group WB accord with the retrieval processes by which S generates his
expectation from this category-cuing analysis. recall. I f the structural information provided
There were relatively few intrusions (20) on by the blocked hierarchies has its main
Trials 1 and 2 for group W B , but all were influence on the retrieval plans for recall,
obviously intrusions of unpresented instances then one should find much less of an effect of
of a presented category, and these intrusions structural information when memory is tested
were "appropriately" placed in the recall by recognition, which largely bypasses the
order. G r o u p W R had 26 intrusions of which retrieval aspects of the task. Accordingly,
25 were clearly within the presented categories. Exp. II was undertaken to see whether the
In theory, these intrusions represent simple large recall difference between the WB and
"false alarms" in recognition of a likely W R conditions would be greatly attenuated
candidate which in fact was not on the list. in recognition tests.
The theory supposes that intrusions decrease
and correct recalls increase with practice
Method
because of two factors: (a) The associations Design and procedure. Each S received two replica-
tions of a two-trial experiment, using the four-
from category to presented instances become
hierarchy sets 1 and 2 from Exp. I. There were eight
stronger, thus causing the latter to be more conditions obtained by crossing Blocked vs. Random
readily and reliably generated as candidates presentation with Recall vs. Recognition over the two
for recall to the former, and (b) the item- trials of each learning task. Using the last-letter
information that aids list discrimination abbreviations L and N for recaU and recognition,
respectively, the four possible test sequences over two
improves, which information may be in the trials are LL, LN, NL, and NN. Subjects in conditions
form either of an estimate of frequency or LN and NL were switched between Lists 1 and 2, as
recency of experience or a trace-strength of were Ss in conditions LL and NN.
the presented items. The four hierarchies of set 1 or set 2 were projected
on four slides in tree form, at 56 see per slide, and then S
EXPERIMENT II either wrote his recall or took a written recognition
test. He did not know during presentation how he
The next experiment investigated whether would be tested. The recognition test used the "Yes-
the large influence of Blocked structural- No" method: A sheet of paper given to S contained
information upon recall could be replicated the 112 list words scrambled amongst 112 distractors,
with a recognition test of memory. A frequent and S was told to check those words which he thought
had been on the slides just studied. Half of the
claim in the recent literature on memory is 112 distractors were unrelated nouns matched in
that recognition measures the amount of Thorndike-Lorge frequency to the list words, while
information stored independently of retrieval the other half of the distractors were conceptually
330 BOWER ET AL.

related words that could have fit into the conceptual the R a n d o m Ss gave 31 of which 29 were
hierarchies presented. The same set of words were categorical.
used for both recognition tests in the NN condition,
Recall o f the eight conceptual hierarchies
except they were listed in a different order on the two
trials. Subjects were run in small groups of two-four, was separately scored, pooling all Ss who were
with test times sufficient for all Ss to finish. A distinct recalling them on either Trial 1 or 2. The
instructional break was made between Lists 1 and 2. recall proportions for the eight hierarchies
Half the Ss were exposed to the Blocked lists and half averaged over Trials 1 and 2 for the Blocked
to the Random lists of Exp. I for both lists and both
condition ( R a n d o m condition in parentheses)
trials. The Ss were 64 Stanford undergraduates from
the introductory psychology course with eight in each were as follows: Body parts, 81(34); plants,
of the eight groups. Pooling across the two lists, there 75(35); minerals, 71(34); animals, 71(33);
are 16 learning protocols for the eight experimental clothing, 71(35); transportation, 69(26);
conditions. instruments, 69(32); and occupations, 61(34).
The variation in recall o f seven o f the eight
Results
r a n d o m hierarchies is too minuscule to give
There were no significant differences be- m u c h o f any correlation with recall o f the
tween first versus second lists nor between sets eight blocked hierarchies ( r = + . 1 7 ) . These
scores m a y be c o m p a r e d to those obtained
TABLE 3
in Exp. I over the same trials but with fewer Ss
PERCENTAGEOF WORDSRECALLEDIN THE
learning each hierarchy there (four Ss instead
Six CONDITIONS
o f 16). F o r blocked presentation, rank orders
Trial o f the recallability o f the eight hierarchies in
the two experiments correlate +.62; for
Condition T1 T~
r a n d o m presentation, the two rank orders
Blocked: correlate +.48. It thus appears that the
L L .61 .89 difference in recallability of the eight hier-
L N* .65 -- archies is small but reasonably consistent
N L -- .88
within a given condition, but that the ordering
Random:
differs consistently for blocked vs. r a n d o m
L L .21 .43
L N .29 -- presentations. We shall look into this
N L -- .39 issue again in Exp. V in which several
normative indices o f the hierarchies are
L N denotes recall on T~ and recognition on T2. obtained.

1 and 2 of the hierarchies, so the data will be TABLE 4

pooled over these variables. The first result RECOGNITIONHITS(H) ANDFALSEALARMS
o f interest is shown in Table 3 which gives ON RELATED(R) AND UNRELATED(UR) DISTRACTORS
percentage correct recalls over Trials 1 and 2
for the Ss that had recall tests. The recall Trial 1 Trial 2
differences between Blocked and R a n d o m Ss Condition H R UR H R UR
are o f similar magnitude to the differences
Blocked:
observed in Exp. I, so that result is replicable. LN .92 .04 0
Further, recall on Trial 2 was approximately N L .88 .05 .01
the same whether Trial 1 was a recall or a N N .80 .07 0 .92 .08 0
recognition test. Recall intrusions were less Random:
than 19/o in all six conditions, but were LN .78 .06 .01
categorical in nature. The Blocked Ss gave N L .61 .06 .03
N N .60 .10 .02 .80 .19 .07
63 intrusions o f which all were categorical;
 HIERARCHICAL RETRIEVAL SCHEMES 331

The relevant results on recognition tests However, an interpretative problem is that

are shown in Table 4 for Trials 1 and 2. The there is no atheoretical way to evaluate the
three columns for each trial give proportions significance of these differences due to Blocked
of (a) checking as "old" an old item (a hit), vs. Random conditions in the recall vs. recog-
(b) falsely checking a conceptually related nition measures. A model relating free recall
distractor, and (c) falsely checking an unrelated to recognition is needed, but none can be
distractor. Hit proportions are based on 112 stated now with any confidence. A simple
observations per S per trial, whereas the false threshold model, which supposes that S recog-
alarm proportions are each based on 56 nizes all those items he could recall plus half
observations per S per trial. of those he could not recall, leads to the
Comparing entries within Table 4, the formula 2 N = I + L , and this gives a fairly
following conclusions are warranted: (a) False good prediction of the average hit rates (Table
alarms were higher on related than on 4) from the average recall results (Table 3).
unrelated distractors; (b) hit rate on Trial 2 However, it gives no prediction of false alarm
was unaffected by type of test on Trial 1, rates, nor their ordering, nor does it provide
but the false alarm rate on Trial 2 .was higher any illuminating analysis of the retrieval
if Trial 1 was also a recognition test; and processes in free recall.
(c) Ss having the Blocked input have better The results of this experiment suggest that
list discrimination than Ss having Random recognition of list membership of a word
input, as indicated by a higher hit rate a n d depends on the structure of the list as well as
a lower false alarm rate. the words that were in it. In particular,
The first conclusion, thatrelated distractors recognition of a g i v e n word apparently
are Checked more often, is hardly surprising depends upon the number and configuration
a n d can be handled theoretically in several of associations converging upon that word
ways. The second conclusion, that false alarms from other list words that have been recently
are highest in the N N conditions, is probably primed. This view accounts for the effect of
a result of using the same distractors on Tests 1 list structure upon hit rate, and it also explains
and 2. Thus, on Test 2 in the N N condition, why related distractors elicit more false
a distractor arouses a familiar sense of having recognitions than unrelated distractors, even
been seen before and is checked more often. in the Random condition. Underwood (1965)
These are ancillary findings. and Anisfield and Knapp (1968) have proposed
The interesting result is the third, that a similar view of recognition memory.
recognition is better l for Blocked than for A n alternative interpretation of these
Random Ss. A t test on recognition hits recognition differences would attribute the
minus false alarms for all Blocked vs. Random effect to "unequal" exposure times to the
Ss on Trial 1 yields t(62)= 5.94, p < .001, individual words in the Blocked vs. Random
and on Trial 2 yields t(62) = 6.69, p < .001. hierarchies. Since only total time was con-
Comparing the averaged Blocked vs. Random trolled, Ss in the Blocked conditions might
results in Table 4 to those in Table 3 shows have used the redundancy or predictability
that the average recall differences were much of the hierarchies to scan rapidly and more
larger than the recognition differences; recall often over all the words. On the other hand,
differences were 38 and 48 ~ on Trials 1 and 2 Ss in the Random condition may have been
compared to recognition differences of 23 much slower in reading the words, and for
and 13 ~o on the same trials. two reasons: (a) Recognition or reading times
We may ask whether the recognition test of words are known to be slower if the word
has attenuated the difference between Blocked sequence is unpredictable, so fewer words
and Random Ss seen on the recall test. would be scanned in a fixed time, and (b) the
 332 BOWER E T AL.

R a n d o m S m i g h t have used u p e x p o s u r e time occur if List 1 consists o f level-4 instances o f

trying to organize o r systematize the r a n d o m c o n c e p t u a l hierarchies, a n d List 2 consists o f
w o r d s he h a d a l r e a d y read, thus reducing the the levels 1-3 s u p e r o r d i n a t e s o f these same
time he c o u l d spend processing later words. hierarchies. I n this case, when S is then a s k e d
A c c o r d i n g to this view, then, the Blocked vs. to recall everything f r o m Lists 1 a n d 2, his
R a n d o m difference in recognition resulted recall o f the s u p e r o r d i n a t e s f r o m List 2 should
f r o m less real exposure to the i n d i v i d u a l w o r d s serve a cuing function in facilitating his recall
in the latter condition. W e are currently o f the List-1 instances.
testing this i n t e r p r e t a t i o n with a m e t h o d o f
presenting the h i e r a r c h y one w o r d at a time, Method
t h u s p e r m i t t i n g c o n t r o l o f S ' s e x p o s u r e time Design andprocedure. There were three experimental
to each word. T h e S sees a c o m p l e t e hier- conditions, a rest control (C), relevant interpolation
archical tree o f unfilled, n o d a l circles with (RI), and irrelevant interpolation (II). All Ss began
lines between nodes, a n d 28 successive slides with two trials on List 1 which consisted of 48 level-4
words from two conceptual hierarchies (e.g., minerals
s h o w in systematic o r d e r one new n o d a l w o r d and animals). These hierarchies were so tailored that
for 2 sec, until e a c h w o r d in the conceptual they each had six groups of four level-4 words. These
h i e r a r c h y has been presented once. I f differen- four level-4 words (subsumed by one level-3 node)
tial exposure time to the individual w o r d s were were presented together as a group for eight seconds by
the f a c t o r d e t e r m i n i n g the results o f EXP. II, a slide projector. In List 1, there were 12 such slides
arranged in a different random order for the two trials.
then the Blocked vs. R a n d o m difference in Subjects gave their free recall by writing the list words,
r e c o g n i t i o n s h o u l d vanish when e x p o s u r e time being permitted 3 min.
is c o n t r o l l e d in this manner. After two trials on List 1, the RI and II Ss learned a
second list while the C Ss read Peanuts cartoons for a
EXPERIMENT I I I time equal to the longest interpolated learning interval.
The interpolated list differed for RI and II Ss. For RI
T h e next e x p e r i m e n t investigates the retro- Ss, the interpolated list consisted of the levels 1, 2 and 3
active effect u p o n free recall o f a first list superordinate words from the two hierarchies relevant
i n d u c e d b y the learning o f a second list. for classifying the level-4 words which S had learned in
Previous experiments b y Tulving a n d T h o r n - his List 1 (e.g., minerals and animals). The relevance of
t o n (1959), a n d P o s t m a n a n d K e p p e l (1967) List-2 to S's List-1 words was not pointed out to him.
For II Ss, the interpolated list consisted of the levels
have shown retroactive decrements in free 1, 2 and 3 superordinate words from two hierarchies
recall. F o r example, in the P o s t m a n a n d which were irrelevant for classifying the prior List-1
K e p p e l study, Ss l e a r n e d List 1, then h a d words (e.g., instruments and occupations.) The two
v a r y i n g n u m b e r s o f trials on List 2, then were List-2 hierarchies for a given S were presented in tree
a s k e d to recall all the w o r d s in b o t h lists. form, one per slide for 18 sec (two seconds per nine
words in each tree), and then S wrote his recall of List 2.
R e c a l l o f List-I items was m a r k e d l y p o o r e r After S reached a criterion of one perfect recall of
t h e greater the n u m b e r o f trials S h a d h a d on List 2, he was then asked to recall (in writing) everything
List 2. T h e a u t h o r s discussed these results in he could from both Lists 1 and 2, starting with List 2
terms o f the u n l e a r n i n g (during trials on if he could. Rest control Ss were also asked to recall
List 2) o f the associations between the experi- List 1 at this time. Four minutes were allowed for this
recall and protocols were collected. At this point, Ss
m e n t a l context stimuli a n d List-1 responses. in groups C and II were given a sheet of paper contain-
These interference o r u n l e a r n i n g effects have ing the relevant superordinate categories for List 1
been p r o d u c e d with u n r e l a t e d w o r d s in Lists 1 (i.e., levels 1-3 arrayed in tree form) and were asked
a n d 2. O u r question here is whether we can to attempt a second recall, of the List-1 words only,
p r o d u c e j u s t the o p p o s i t e effect, retroactive since "the words on this sheet may help you to do
better in your recall" of List 1. Three minutes were
facilitation, if the two w o r d lists fit into the allowed for this cued recall by C and II Ss.
s a m e c o n c e p t u a l hierarchies. I n p a r t i c u l a r , After the procedures described above were com-
retroactive facilitation o f List-1 recall m i g h t pleted, the entire experiment was replicated with
 HIERARCHICALRETRIEVALSCHEMES 333

completely different word hierarchies partitioned to recall to the corresponding recall scores on
serve as List 1 and List 2 but with S assigned to the Trial 2 (row 1), we see that there was practi-
same treatment condition as before. The order in
cally no change for the C and I! Ss, but a large
which the various word hierarchies were used (first or
second replication) was counter-balanced over Ss. increase for RI Ss. The increase in recall for
The Ss were 33 undergraduates fulfilling a service RI Ss is highly significant: t(21)=7.95,
requirement for an introductory psychology course. p < .001. On this later test, the gain in recall
They were assigned 10 to group C, 11 to group RI, relative to what RI Ss could have gained
and 12 to group II. The Ss in a given condition were
run in small groups of two to five at a time. was 53 Yoo.
A similarly significant increase in List-1
Results recall occurred for groups C and II when the
There was a significant improvement in relevant superordinates hierarchies were given
Ss' performance from replication-1 to repli- as recall probes on their last test (row 3 of
cation-2. Considering List-1 performance Table 5). The II Ss gained 42 ~ and the C Ss
within each replication, recall proportions gained 39 ~o of the responses they could have
gained over Trial 2 with List 1. A comparison
averaged 52 and 68 ~ on Trials 1 and 2 of
of the probe recall scores of C and II Ss with
replication-l, but were 59 and 79 ~o on Trials 1
and 2 of replication-2. This general practice that of RI Ss on their "recall everything" trial
yielded no significant differences, F(2, 2 8 ) =
TABLE 5 2.07, p > .10.
RECALL OF THE 48 LEVEL-FouRWORDSOFLIST1
The simplest explanation of these results
would,attribute the List-1 recall increments to
List 2 the cuing function of the relevant super-
Rest ordinate words. That is, on Trial 2 with List 1,
Control Irrelevant Relevant
whole groups of level-4 words might not have
1. Trial 2 of List I 35.3 34.5 35.6 been recalled because S had not yet developed
2. Recall all 34.9 34.5 42.2 a scheme for cuing his recall of these. However,
3. Relevant probe 40.3 40.1 -- when the relevant superordinate categories
are available on S's recall sheet, either by E's
effect, however, is orthogonal to the major provision or by S recalling them, they lead
factors of interest in this experiment, so him to think about practically all the groups
results will be pooled over replications for the of level-4 words and to recall some words of
analyses which follow. these groups. This account is bolstered by
There were no significant differences among analysis of the cluster-recall data. W e scored
the three groups in their acquisition of List 1. a level-4 duster as recalled if at least one of
The average List-1 recall scores on Trial 2 the four words in that group was recalled.
(the last acquisition trial) are shown in the For the clusters that were recalled, we also
first row of Table 5 and are obviously similar. computed the mean words recalled per cluster.
Learning of the List-2 blocked hierarchies by These measures for Trial 2 of List 1 and for the
RI and II Ss was very rapid, with the median S final recall trial are shown in Table 6 for the
recalling his List-2 perfectly on Trial 2. These three treatment groups. The groups differ
Ss also recalled their List-2 perfectly on the relatively little in the percentage of clusters
"recall everything" test which immediately recalled or in the mean words per recalled
followed their criterion trial. cluster. The effect of hierarchical cuing
The results of major interest are the List-1 provided by S or E was to increase the
recall scores for the "recall everything" test percentage of clusters recalled but not the
given after List-2 learning; these are recorded mean words per recalled cluster. Within each
in the second row of Table 5. Comparing this group separately, the increase in proportion
334 BOWER JET A L .

TABLE 6 Ss. Thus, this one implication of the media-

PROPORTION OF CLUSTERS RECALLED AND MEAN tional interpretation was not supported by the
WORDS PER RECALLED CLUSTERON TRIAL 2 AND FINAL evidence.
RECALL TRIAL
EXPERIMENT I V
Proportion of Words per cluster
dusters (of4) The first three experiments have examined
the influence of structured input upon memory
Group Trial 2 Final Trial 2 Final
for conceptual hierarchies. The relationship
Relevant .84 .98 3.52 3.38 between successive nodal words in a con-
Irrelevant .89 .99 3.20 3.35 ceptual tree is roughly one of"dass inclusion,"
Rest Control .86 .98 3.40 3.42 or "instance of a superordinate category."
In terms of associationistic psychology,
however, these class inclusion relationships
of dusters recalled is significant (t = 8.00, are only one of many bases upon which two
12.20, 5.80 for groups RI, II and C, respec- words could become associated. Taking this
tively); but for no group was there a significant broader view, then, a conceptual hierarchy is
change in the mean words per recalled cluster. just a specific kind of associative hierarchy or
We have attributed the retroactive facilita- network, one for which the associations
tion of recall in group RI to the simple cuing between nodal words have approximately the
effect of having the superordinates available same basis, namely, class inclusion. But this
at the time of recall. This is in accord with the broader view immediately leads one to con-
results of Tulving and Pearlstone (1966) and sider other kinds of associative trees--
Wood (1967b), who found more words specifically, trees in which successive nodes
recalled if Ss were provided with relevant are associated with one another, but in which
category cues at the recall test. An alternative the basis for the association may vary freely
interpretation might attribute the RI result over different nodal pairs in the same tree.
to strengthening of List-1 words while S was The tree in Fig. 2, for cheese, will illustrate
studying the relevant List-2 superordinates. our points about associative hierarchies. We
That is, while viewing the relevant super- constructed such trees by starting with a level- !
ordinates, S might notice their relationship to nodal word, and recording three different but
List-1 responses and implicitly rehearse the strong verbal associates to it as level-2 nodes.
List-1 responses, associating them with the Each of these second-level nodes was then
superordinate context of List 2. This view considered in turn and to each was recorded
would provide a "mediational" interpretation two different associates as level-3 nodes;
of our RI result similar to that provided for finally two associates were recorded below
the retroactive facilitation observed with the each level-3 node, giving a four-level hier-
A-B, A-B' paradigm in paired-associates archy of 22 words. We have constructed
learning, where A is the cue, and B and B' eight of these associative trees. They are based
are associatively related response terms (cf. only on intuitive judgments of pairwise
Postman, 1961). An implication of this inter- associations, since published association
pretation, however, is that final recall for RI norms are too restricted to provide guidelines
Ss should have exceeded probed recall for the for constructing such depth hierarchies. There
C and II Ss, since the RI Ss allegedly have had is obviously nothing unique or special about
the benefit of extra rehearsal on List-1 words such associative hierarchies. There are doubt-
during their List-2 learning. In fact, howeveL less thousands that could be constructed
there was no reliable difference between final starting with almost any content word in
recall for RI Ss and probed recall for C and II node-l, with variation in the number and
 HIERARCHICAL RETRIEVAL SCHEMES 335

l
LEVEL
~CHEES

Fro. 2, The associative hierarchy for "cheese."

d e p t h o f branchings. T h e associative base for words followed by a blank recall sheet. Instructions
the hierarchies in Fig. 2 is n o t easily specified were given orally by E.
There were four types of booklets distributed each
since the type o f a s s o c i a t i o n relating node-n
day, each to one-fourth of the class. One type presented
to n o d e - n + l varies considerably. A s s o c i a - two hierarchies in an organized (blocked) tree form,
tions between successive pairs are "intuitively and the second used these same words in two trees,
sensible," but, like free-associative chains, the but with total scrambling of the words across associa-
c o n n e c t i o n s a m o n g m o r e d i s t a n t elements are tive hierarchies and levels. The third and fourth types
were the same as the first and second types, except
sensible only b y v i r t u e o f the intervening
that the order of presentation of the two hierarchies
pairwise associations in the chain. on each trial was reversed. On the first day, the wish
G i v e n these associative hierarchies, one can and hammer hierarchies were used, and on the second
a g a i n a s k the same q u e s t i o n as in Exp. I ; day the salt and cheese hierarchies were used. Thus,
n a m e l y , is free recall facilitated b y presenting the second day was a replication of the first, with new
words, except that each S received the condition
t h e m in a structurally b l o c k e d f a s h i o n as
opposite to what he had received on Day 1 (i.e.,
o p p o s e d to a r a n d o m f a s h i o n ? A c c o r d i n g l y , organized-random order or vice-versa).
a simple c o m p a r i s o n was carried o u t between The hierarchies of 22 words were printed in vertical
c o m p l e t e p r e s e n t a t i o n o f these b l o c k e d hier- tree form (with no circles or lines), one per study sheet.
archies a n d c o m p l e t e p r e s e n t a t i o n o f the Each was studied for 45 see as timed by a stop-watch
by E, with instructions to turn the page to the next
s a m e w o r d s m i x e d u p r a n d o m l y . The two
study sheet. After studying both hierarchies, at a
c o n d i t i o n s are similar to the W B a n d W R signal S turned to a blank recall sheet and tried to write
c o n d i t i o n s of'Exp. I. all the words he had studied, in any order. Recall time
was 3½ min; its termination was indicated by E
Method and the next study trial began.
One S was dropped because he was not available
The Ss were 44 students in a high-school plane
the second day, and one because she obviously
geometry class made available to us in a local public
misunderstood the instructions on the first day. This
school system. The students were mainly tenth graders,
left 42 usable Ss with two learning protocols each.
with a few from grade 11. They were tested for 30 min
on two consecutive days, having four input-output
trials on different hierarchies on each of the two days. Results
Their-regnlar teachercooperated as one of the Es and
in proctoring their performance3 There were insignificant differences between
Design and procedure. The Ss received mimeo- recall on D a y s 1 a n d 2 within a given condi-
graphed booklets of the study materials and recall tion, so all p r o t o c o l s within a c o n d i t i o n were
sheets. Each booklet consisted of 12 pages, four sets
of three pages, consisting of two study pages with list p o o l e d i g n o r i n g this D a y s factor. T h e m e a n
w o r d s recalled (out o f 44) for the B l o c k e d
2 Our thanks to Mrs. Sharon Lesgold, the teacher, c o n d i t i o n on Trials 1--4 were 23, 34, 39, a n d 41 ;
for helping us with this experiment. the c o m p a r a b l e m e a n s for the R a n d o m
336 BOWER ET AL.

condition were 16, 23, 28, and 33. Mean Exp. I. There were, of course, many procedural
recall in the Blocked condition significantly differences between the two experiments--
exceeded that of the Random condition on different S-populations, different exposure
each trial (all p's < .01). Each S learned both and recall methods, etc. Ignoring these
a Blocked and a Random list; 40 of the 42 Ss procedural variations, however, it appears that
recalled better with their Blocked list. the blocked associative hierarchies are more
Recall of the four different associative difficult than the blocked conceptual hier-
hierarchies was quite homogeneous. Percent- archies, whereas the two random conditions
ages recalled averaged over the four trials (conceptual vs. associative words) would
were as follows for the Blocked (and Random) appear to be about comparable in difficulty.
hierarchies: Wish, 81(61); hammer, 79(59); A direct experimental comparison of the two
cheese, 77(59); and salt, 74(56). The differences types of hierarchies is being done to check
in recallability of the four hierarchies are small on this ordering.
and insignificant within both the Blocked and The main conclusion to be gathered from
Random conditions. the results of this experiment is that blocking
An analysis of ordered recall similar to that of associative hierarchies considerably im-
of Exp. I (which used oral recall) was not done proves recall over a condition of random
because of equivocation in deciding in which input of the same words. What is apparent to
order many Ss had written down the words Blocked Ss but not to Random Ss is that many
they recalled. Some Ss in the Blocked condi- word pairs are associated and moreover that
tion wrote their recall in the tree form they many lengthy associative chains exist in the
had studied", but the temporal order in which word lists. The "structural principle" for the
they had recalled the nodes could not be Blocked lists is that of recursive rewriting
determined. It was possible to score the according to associative transitions, going
protocols for joint recall of the level-3 and its from each level-n node to its several level-n+l
level-4 words. Considering Trial 1 with the nodes. Subjects notice and comment upon this
Blocked list, the probability of recalling a construction principle, and they obviously use
level-4 word was .86 when its level-3 word was it for reconstructing the list from memory.
recalled, but only .36 when its level-3 word Again however, the recursive principle is not
was not recalled. The comparable conditional allowed to run unchecked, since it alone
recall proportions for the same words in the would produce many associative intrusions of
Random lists were .39 and .32, respectively. nonlist words. Rather the candidates produced
A similar pattern held for recall of level-3 by the associative recursion must be checked
words conditional upon recall of level-2 words. for list-membership information before they
This pattern of results again points to the are overtly recalled. The few intrusions that
cuing role of the level-n node in the recall of its do occur for the Blocked Ss, however, appear
level-n+l words. This cuing effect is much predominantly to be "false recognitions" of
more potent with the Blocked than with the candidates generated by the associative recur-
Random hierarchies presumably because the sion principle.
n-to-n+l association is immediately available
for rehearsal in the Blocked list, but is buried
among many competing responses in the EXPERIMENT V
Random list. A possible objection to our experiments so
While the recall differences between the far might be that they merely show that
Blocked and Random conditions are large Blocked Ss have learned the principle of list
here, they are not as dramatic as the differences construction, and that knowledge of this
obtained with the conceptual hierarchies in principle is sufficient for them to produce
 HIERARCHICAL RETRIEVAL SCHEMES 337

most of the words in the hierarchies without Method

having any further specific information about The Ss, ten Stanford undergraduate volunteers
the input list. For example, Ss told to write enrolled in Summer classes, were run individually.
the numbers from 1 to 112 would produce After general instruction illustrated with one example,
S began work with the four associative hierarchies
more than Ss shown these same numbers on
from Exp. IV, and then did the eight conceptual
many slides in a randomly scrambled order hierarchies from Exp. I. The order of working on the
and told to recall. The difference in "recall" hierarchies within each set was randomized over Ss.
would be utterly trivial in this instance because All work was completed on one hierarchy before S
the list-construction principle suffices com- went on to the next hierarchy. Work on a hierarchy
began by giving S a slip of paper with the level-1 word
pletely to generate all list items and no nonlist
printed in a box, and with 2 (or 3) lines radiating down
items. Are our Blocked vs. Random compari- to two (or three) empty boxes at level-2. The S wrote
sons in the prior experiments simply showing in what he thought were the appropriate words. He
this same trivial outcome in a slightly veiled was then shown a second slip of paper with the "right"
disguise ? level-I and level-2 words, with lines radiating down to
the appropriate number of level-3 empty boxes which
We think not. The difference between the
he was to fill. After filling these, he was given a p a p e r
iaumbers case and our hierarchies is in the with the level 1-3 words and filled in the appropriate
range of acceptable alternatives that could be number of boxes for level-4 words. The recursive
generated by the list-construction principle in "resetting" of S back on the "right" track gave him
the two cases. In the former case, the principle feedback on how well he was doing. Occasionally S
could not think of enough appropriate words to fill the
by itself generates all list items and no others;
indicated number of empty boxes, and he was permitted
in the latter case, the principle generates only to leave such spaces blank. Testing on the complete
a few list items and very many nonlist items series required 30-45 min depending on S's speed
(intrusions?). We shall return later to a of generating appropriate words. Because of an
discussion of the psychological distinction experimental error, responses of one S were lost on
level-2 responses to one conceptual hierarchy
between these two cases; but for our immediate
and level-3 responses to another conceptual hier-
purpose it is best to collect some evidence for archy.
our claim that the word-hierarchies we have Scoring. For notation, let k denote the total number
used here are quite open-ended and non- of responses required of S at a given level of the tree;
exhaustive. k is the product of the number of nodes at that level
times the number of responses required at each node.
Experiment V essentially collected associa-
With ten Ss, we shall have a distribution of 10k re-
tion norms on these hierarchies from a group sponses (minus a few omissions) at each level of each
of Ss. The S was given a general description hierarchy. A variety of indices could be computed
of the list-construction principle and then from such associative distributions. We shall report
tried to generate the word hierarchies. His two. The first measure will be the proportion of the 10 k
responses which match the "correct" words at that
generation was always set back on the "right
level of the hierarchy. This will be relevant for inter-
track" when he erred. Given the complete preting S's ability to generate the particular hier-
tree up through the level-n nodes, S was to archies which E had constructed.
generate the appropriate number of level-n+l The second measure is one of dispersion, or con-
words under each node. After this, the versely, how much Ss agreed on their responses at
a given node. A convenient measure of agreement may
"right" level-n+l words were added to the be obtained as follows: the responses at a given level
tree, and S was asked to produce the appro- of a tree are rank ordered according to their frequency
priate number of level-n+2 words. Both the (with omissions considered as unique responses), and
conceptual and the associative hierarchies of then we see how far down in the rank ordering one has
Exps. ! and IV were used. The question was to go to accumulate 50 % of the response distribution.
The more agreement of Ss' responses, the lower will
simply, how well will Ss do in generating such be the rank of the median response. The rank of the
trees knowing only the principle by which the median will depend on k, unfortunately, so this score
trees were constructed ? will not be comparable across levels unless it is trans-
338 BOWER E T AL.

formed to a common scale. This is achieved in the average o f the proportions correct at each
Agreement Index (,41), level, with weighting coefficients equalling the
Nk proportions o f words in the hierarchy contri-
----X
2 N 2x buted by the words at a given level. The column
AI=Nk k N-1 k(N-1) "
labeled N R reports the total n u m b e r o f
2 2
different responses (types) given for the entire
where N is the number of Ss, k is the total responses
per S at a given level, and x is the obtained rank of the hierarchy by the 10 Ss, c o u n t i n g omissions
median word at that level. The A I index has the as unique responses (i.e., considering them to
following rationale: (a) If all N k responses differ have been filled in by r a n d o m selection f r o m
(i.e., no agreement), then the median rank, x, will be a dictionary). Excluding the level-1 name o f
Nk]2 and A I will equal 0; and (b) if all N Ss agree in the hierarchy which S was given at the start,
giving the same k responses, then the median rank,
the target associative hierarchies had 21
x, will be k[2, and A I will equal 1. Therefore, A I
scales from 1 for complete agreement down to 0 for different words, and the target conceptual
complete disagreement. hierarchies had f r o m 25 to 30 different words,
averaging 27. The N R measure thus gives a
Results r o u g h index o f associative variety relative to
The values o f these indices for the four the target hierarchies, The final column,
associative hierarchies o f Exp. IV and for the labeled Recall, gives the average recallability
six conceptual hierarchies o f Exp. I are shown o f these hierarchies when they were presented
in Table 7. The total p r o p o r t i o n correct for in blocked form, pooling all four trials f r o m
the overall hierarchy is obtained by a weighted Exp. IV for the associative hierarchies and the

TABLE 7
MEANPROPORTIONCORRECT(/7), AGREEMENTINDEX(AI), TOTALNUMBEROF
DIFFERENTRESPONSES(NR), ANDRECALLFORTIlE 12 HIERARCHIES

Level 2 Level 3 Level 4 Total

Associative
hierarchy p AI p AI p AI p NR Recall

Salt .35 .74 .03 .61 .27 .83 .21 115 .74
Cheese .10 .44 .20 .78 .14 .58 .18 134 .77
Hammer .43 .89 .20 .78 .09 .64 .18 126 .79
Wish .37 .81 .20 .80 .14 .72 .21 123 .81
Average .31 .72 .16 .74 .16 .69 .19 124 .78

Conceptual
hierarchy p AI p AI p A1 p NR Recall

Instruments .30 .67 .23 .76 .42 .93 .37 145 .71
Minerals 0 .56 .02 .64 .51 .95 .39 120 .74
Body .23 .48 .55 .96 .48 .94 .47 103 .81
Plant .06 .50 .07 .63 .29 .76 .24 160 .77
Animal 0 .48 .22 .72 .27 .90 .24 147 .72
Transportation .70 .96 .13 .81 ;21 .86 .27 117 .72
Occupation .20 .63 .16 .78 .34 .90 .30 140 .64
Clothing .20 .39 .23 .85 .36 .96 .35 95 .71
Average .21 .58 .20 .77 .36 .90 .33 128 .73
 HIERARCHICAL RETRIEVAL SCHEMES 339

first two trials pooling Exps. I and "II for the situation. A similar statement applies for
conceptual hierarchies. relating generation probability of the associa-
The scores in Table 7 vary widely and fall tive hierarchies to their recallability from a
into few discernible patterns. For the con- random list; the range of variation in recall
ceptual hierarchies, the proportion Correct and from the randomly presented lists (Exp. IV)
the agreement indices increase at deeper levels was only 5 ~ , which is too small to consider
of the tree. For the associative hierarchies, correlating that variable with generation
the average proportion correct decreases probability.
somewhat at deeper levels of the tree, but the Turning to the conceptual hierarchies, the
average agreement index remains constant situation there is somewhat better for investi-
across levels. These patterns of changes in AI gating the relation between generation proba-
for the conceptual vs. associative hierarchies bility and recall, since there was larger
are understandable in terms of the increasing variation in both factors. However, the
constraints upon responses imposed by the covariation in generation probability and
greater contextual information at the deeper recallability is not at all strong in this case.
levels of the conceptual but not the associative The body hierarchy was highest on both
hierarchies. The overall proportion correct is variables, but there were several large dis-
higher for the conceptual hierarchies, but this crepancies in rankings of other hierarchies on
is due to the preponderance of level-4 words the two variables. If the eight hierdrchies are
which are more constrained and better guessed rank ordered on both variables, the correlation
in the conceptual hierarchies. of the rank orders is only +. 14, indicating avery
The N R or associative variety measure weak relationship between the two variables.
reveals the open-ended or nonexhaustive The Pearson correlation coefficient between
character of all these hierarchies. The number the two variables is +.40, but this is well below
of different responses given by ten Ss is the value of .62 required to reject the null
consistently 4-7 times more than the number hypothesis of zero correlation at the 5 ~o level.
of words in the hierarchies, with the propor- The generation probabilities of the conceptual
tions in excess of the target words being hierarchies were also correlated with the
somewhat higher for the associative than for recallability of these words when they were
the conceptual hierarchies. Again, this seems presented in random order (condition WR
attributable in part to the greater constraints in Exp. I and II). Here, too, the correlations
in the conceptual hierarchies imposed by the were low and insignificant: Rank-order
restricted inclusion or part-whole relation p = +.24; Pearson r = +.40 but seven of the
between successive nodes. eight hierarchies differed by only 3~o in
Finally, we examine the relation between average recallability. These results show that
recallability of the hierarchies and some recallability of the individual hierarchies is
overall index of naive Ss' ability to generate the not predictable from the probabilities with
different hierarchies. The mean recall probabi- which naive Ss can generate the target hier-
lities for the blocked hierarchies are shown in archies. In a similar vein, recallability of the
the final column of Table 7. For the associative conceptual hierarchies correlates very poorly
hierarchies, there is only very small variation (p = +.15) with the N R index of associative
in recallability and in overall generation variety. Possibly other indices of hierarchic
probability for the four hierarchies. Because integration or association could be composed
the range exhibited in both variables is small which will correlate better with recallability of
relative to the probable error variance of these these 12 hierarchies; but the clear failings of
measures, no systematic correlation of the two these obvious-variables prognosticate little
variables could possibly be shown in this success for such a search.
340 BOWER E T AL.

One might entertain the hypothesis that the explaining the effect of Blocked vs. Random
observed differences in free recall between presentation; namely, the Blocked spatial
the Blocked vs. Random input conditions in array enables S to directly strengthen parti-
the prior experiments simply reflects the differ- cular category-instance associations and it
ential guessing of the words when S does or also provides him with a systematic (hier-
does not know the principle of hierarchic archic) plan for cuing these candidates for
construction. For example, in Exp. IV with the recall, which candidates are recalled only if
associative hierarchies, the Blocked vs. they pass a recognition test for list member-
Random difference in recall, averaged over the ship.
four trials, Was .19, which is exactly the
average generation probability for these DISCUSSION
hierarchies in Table 7. However, there are The message of these studies is simple: If S
several deficiencies in this account of the can discover or learn a simple rule or principle
Blocked vs. Random differences in recall. which characterizes the items on a list and
First, with the conceptual hierarchies, the which relates them to one another, then he
average generation probability (.33) con- uses that rule as a retrieval plan in reconstruct-
siderably underestimates the recall advantage ing the items from memory, with a consequent
of the Blocked list on Trials 1 and 2 (average improvement in his performance. The prin-
difference ca..48). Second, it is not true that ciple characterizing a hierarchy is that of
Random Ss would never guess words on these recursive rewriting by associative transitions.
hierarchies, since the general categorical For our conceptual hierarchies, the associative
nature of the word sets was still fairly obvious transitions are primarily of one restricted
to Ss receiving the random presentations, as type, namely, the relationship of class inclusion
shown by the categorical nature of the intru- or part-whole. But Exp. IV showed facilitation
sions by these Ss; so this factor would tend to even when the base-type of associative transi-
reduce the guessing differential below the. 19 tion varied widely over the set of words. The
or .33 figures of Table 7. Third, if Ss were potent effect of such retrieval plans on free
really guessing words from the conceptual recall was illustrated in Exp. I, where Ss were
hierarchies, then one should find a very large recalling 112 words perfectly after two or three
number of categorical intrusions, since the NR input trials, which may be compared with any
measure in Table 7 shows that these hierarchies number of other reports in the literature
are very open-ended. In fact, however, recall showing much poorer recall of shorter, less
intrusions never exceeded 1 ~ in either the organized lists.
Blocked or Random condition on any trial. A question to which we must return is
The upshot of these considerations is that whether this result is trivial, whether it is
the differential recall produced by the Blocked analogous in some obscure way to asking S
vs. Random presentations is not consistent to "recall" the numbers from 1 to 112 To
with a differential guessing probability exhi- answer this objection adequately would
bited by Ss who do vs. Ss who do not know the require a fairly detailed analysis of the opera-
principle of hierarchic construction. Nor is the tions involved in "memory experiments" (free
recallability of a given hierarchy, with either recall in particular), the involvement of
Blocked or Random presentation, related in response-generation rules in such experiments,
any significant degree to the average probabi- and the difference between "remembering"
lity with which a naive S can generate the items from a presented set versus generating
words in that hierarchy from the construction those items given only knowledge of a rule
principle. We are therefore left with the characterizing the set. These are difficult
theoretical reconstruction given earlier for questions, but we would nominate them as
 HIERARCHICAL RETRIEVAL SCHEMES 341

extremely important questions to answer if one Miller's experiment, the list subset was not
wants to know what is involved in the especially restricted (different initial letters
"remembering" seen in free recall. and strings of differing lengths) and S did not
The observable behavioral difference be- know the principle of list construction (the
tween remembering a set of items and rewrite rules) at the start of the task. However,
generating a list according to a principle is in there can be little doubt that the recall
the nature and extent of intrusion errors. Of improvement observed with practice was in
course, whether a particular candidate gene- part due to the S learning some or all of the
rated by the rule is or is not an intrusion rewrite rules exemplified by the list strings
depends on how well the rule generates only (see also Smith, 1966).
those items on the list and no others. To the The experiments by Whitman and Garner
extent that the rule fails to discriminate well (1962) (also Garner and Whitman, 1965) are
between list and nonlist candidates, to that relevant because their Ss in principle knew how
extent specific item-memory has to be used to generate all possible items of the population
along with the rule to make it an adequate from which a subset was to be recalled. A
recall device. And it has been conjectured here typical task might expose S to eight of the
that this item-memory is essentially "occur- possible 16 geometric figures that could be
rence" recognition, mediated by some kind of composed by combining four binary attributes
recency or contextual frequency tag stored (e.g., one or two, large or small, red or blue,
along with a lexical item in S's long-term circles or squares). The principle for con-
semantic memory. The evidence collected in structing the population of patterns was simple
Exp. V tends to discredit the view that Blocked enough that most Ss could probably generate
Ss are "merely guessing" from their long-term all 16 patterns in the population after only a
semantic memory in a manner consistent with few exposures. The problem for the free-recaU
their knowledge of the list-construction S in such experiments is not one involving an
principle. inability to generate items on the presented list,
It is of interest to compare the present but rather one of suppressing intrusions, of
experiments to those of Whitman and Garner inferring a restricting rule which will generate
(1962) and Miller (1958) which also involved only those patterns presented and no others
free recall of rule-generated lists. In Miller's of the population. Whitman and Garner
experiment, Ss saw many strings of 3-5 letters showed that the difficulty of free recall in this
and then tried to recall them. For one list, situation depended upon the complexity of the
the strings were a subset of those produced by restrictive rule characterizing the subset of
a simple Markov generator, which is a set of presented patterns. For instance, if the
left-to-right rewriting rules for a finite presented subset were to occupy an elementary
vocabulary. For example, the three rewrite partition of the population (e.g., the eight red
rules A ~ B, B ~ (C or B), and C ~ (A or C), figures), then free recall would be extremely
will generate strings like AB, BBC, CCAB, easy. Whitman and Garner also point out the
BCCABBC but not CB, BA, or BAC. Miller basic similarity between such recall experi-
found that a list of such rule-generated strings merits and concept learning experiments, if One
was more easily learned and recalled than was identifies the presented subset for free recall
a list of random strings of comparable length with the list of "positive instances" of a
with the same letters. If the list subset had had concept experiment. A rule for generating
a simple structure (e.g., all permissible strings free recall of the presented list would have as
of length 4 beginning with C), then knowledge an essential ingredient the concept charac-
of the rewrite rules would have permitted S terizing the members of that subset of the
to "recall" all the presented strings. But in total population.
342 BOWER ET AL.

The relevance of these studies by Miller and is that it should be a familiar or easily learned,
Whitman and Garner to our own experiments summary abbreviation suggesting or leading
is relatively direct. In each experiment, the to the words on the list. The word "plan"
presented items were characterized by some need not suggest any elaborate mentalistic or
structural principle which could be used to cognitive reconstruction of behavior: Millen-
generate these items, and learning was largely sen (1967) and Suppes (1968) have shown how
a matter of discovering the rules and then hierarchical TOTE units (cf. Miller, Galanter,
recognizing the presented from nonpresented and Pribram, 1960) can be analyzed in terms
candidates that could be generated by the rules of conditional implicit stimulus-response
applied to the base vocabulary. In each connections, so this language is noncommittal
experiment, of course, the difficulty of recall with respect to S-R analyses of recall.
depended upon the structural characteristics • The evidence to date, collected when S is
of the whole list and was not some simple given a systematic retrieval plan or mnemonic,
combination of the difficulty of the individual tells us that such plans are sufficient to produce
items considered in isolation. One might say very high levels of recall. The evidence does
that the availability or recallability of a given not yet prove that such a plan is necessary or
item depended upon "wholist" properties of required for S to produce high levels of recall,
the list in which it was embedded. but this hypothesis is beginning to look
According to the viewpoint espoused here, increasingly attractive.
free recall is mediated by S using a retrieval
plan for cuing or generating plausible candi-
dates to recall, and an "executive editor" which
REFERENCES
checks these candidates for recency recogni-
tion before overtly recalling them. Dale (1967), ANISFIELD, M., AND KNAPP, M. Association, synony-
Kintsch (1968) and others have proposed mity, and directionality in false recognition.
Journal of Experimental Psychology, 1968, 77,
similar views. The last few input items in the 171-179.
echo box of short-term memory may be BOWER, G. H. A multicomponent theory of the
recalled directly without mediated cuing, but memory trace. In K. W. Spence and J. T. Spence
the hypothesis supposes that most of the (Eds.) The psychology of learning and motivation.
other items in free recall would be generated Vol. 1. New York: Academic Press, 1967.
CORER, C. N., BRtSCE,D. R., AND REICHER, G. M.
by a cuing system. The cuing system could be Clustering in free recall as a function of certain
a set of rules, some structural information methodological variables. Journalof Experimental
about the composition of the list, an alphabetic Psychology, 1966, 71, 858-866.
scheme (Earhard, 1967), or a pegword system COaEN, B. H. Some-or-none characteristics of coding.
(Wood, 1967a). Last but not least the effective Journal of Verbal Learning & Verbal Behavior,
1966, 5, 182-187.
cue for recall of a given word may be prior COHENs B. H., AND BOUSFIELD,W. A. The effects of a
recall of other words. But even this has to dual-level stimulus-word list on the occurrence of
regress eventually back to an implicit cuing clustering in recall. Journal of GeneralPsyehology,
system. The improvement in free recall over 1956, 55, 51-58.
multiple trials would result, on this view, from DALE, H. C. A. Familiarity and free recall. Quarterly
Journal of Experimental Psychology, 1967, 19,
several factors: (a) more discriminating 103-108.
occurrence-information being stored for list EARSARD,M. Cued recall and free recall as a function
words, (b) increasing integration of subjective of the number of items per cue. Journal of Verbal
dusters or groups of words which S treats as Learning & VerbalBehavior, 1967, 6, 257-263.
GARNER,W. T., ANDWHITMAN,J. R. Form and amount
units, and (c) development of a more adequate
of internal structure as factors in free-recall
retrieval plan for S to cue recall of his subjec- learning of nonsense words. Journal of Verbal
tive units. The characteristic of the cuing plan Learning & VerbalBehavior, 1965, 4, 257-266.
 HIERARCHICALRETRIEVALSCHEMES 343

KINTSCH,W. Recognition and free recall of organized POSTMAN,L., ANO KEPPEL, G. Retroactive inhibition
lists. Journal of Experimental Psychology, 1968, in free recall. Journal of Experimental Psychology,
78, 481-487. 1967, 74, 203-211.
MANDLER, G. Organization and memory. In K. W. Sear,, K. H. Grammatical intrusions in the free recall
Spence and J. T. Spence (Eds.) The psychology of of structural letter pairs. Journalof VerbalLearning
learning and motivation, Vol. 1. New York: & VerbalBehavior, 1966, 5, 447--454.
Academic Press, 1967. SLrpPES,P. Stimulus-response theory of finite automata.
MANDLER, G. Association and organization: Facts, Technical Report 133, Psychology Series, Institute
fancies, and theories. In T. R. Dixon and D. L. for Mathematical Studies in the Social Sciences,
Horton (Eds.) Verbal behavior and general Stanford University, Stanford, Cal., 1968.
behavior theory. Englewood Cliffs, N.J.: Prentice- TULVINO, E., AND PEARLSTOr,m,Z. Availability versus
Hall, Inc., 1968. accessibility of information in memory for words.
MILLENSON,J. R. All isomorphism between stimulus- Journal of Verbal Learning & Verbal Behavior,
response notation and information-processing 1966, 5, 381-391.
flow~ diagrams. Psychological Record, 1967, 17, TULVING, E., AND THORNTON,G. B. Interaction be-
305-319. tween proaction and retroaction in short-term
MILLER, G. A. The free recall of redundant strings of retention. Canadian Journal of Psychology, 1959,
letters. Journal of Experimental Psychology, 1958, 13, 255-265.
56, 485-491. UNDERWOOD, B. J. False recognition by implicit
MILLER, G. A., GALANTER,E., ANDPRIBRAM,K. Plans verbal responses. Journal of Experimental Psycho-
and the structure of behavior. New York: Holt, logy, 1965, 70, 122-129.
Rinehart & Winston, Inc., 1960. WHITMAN, J. R., AND GARNER, W. R. Free recall
MILLER, G. A., AND SELFRIDGE,J. Verbal context and learning of visual ~gures as a function of form of
the recall of meaningful material. American internal structure. Journal of Experimental
Journal of Psychology, 1950, 63, 176-185. Psychology, 1962, 64, 558-564.
PAIVlO, A., YUILLE, J. C., A~NO MADIGAN, S. A. WOOD, G. Mnemonic systems in recall. Journal of
Concreteness, imagery, and meaningfulness values Educational Psychology Monograph Supplement,
for 925 nouns. Journalof ExperimentalPsychology 1967a, 58, no. 6, Part 2, 1-27.
Monograph Supplement, 1968, 76, No. 1, Part 2, WooD, G. Category names as cues for the recall of
1-25. category instances. Psychonomic Science, 1967b,
PARKS,T. E. Signal detectability theory of recognition 9, 323-324.
memory performance. Psychological Review, WooD, G., AND UNDERWOOO,B. J. Implicit responses
1966, 73, 44-58. and conceptual similarity. Journal of Verbal
POSTMAN,L. The present status of interference theory. Learning & VerbalBehavior, 1967, 6, 1-10.
In C. N. Cofer (Ed.) Verbal learning and verbal
behavior. New York" McGraw-Hill, 1961. (Received November 8,'1968)

View publication stats