Grouping Operations in Free RecallZ

Transcription

1 JOURNAL OF VERBAL LEARNING AND VERBAL BEHAVIOR 8, (1969) Grouping Operations in Free RecallZ GORDON H. BOWER, ALAN M. LESGOLD, AND DAVID TIEMAN Stanford University, Stanford, California In free recall, Ss search for stable groupings of the list words, which groups become functional recall units. The experiments reported show that recall suffers if S is forced to adopt groupings differing from those used previously. Although such regroupings disrupt retrieval processes in recall, they do not reduce recognition memory for the list words, presumably because recognition depends on "occurrence information" which is independent of interitem associations utilized in retrieval. Another experiment showed that S's recall is facilitated by E explicitly aggregating S's functional recall units into larger groups over trials. A final experiment showed that recall was facilitated by arranging for structural linkages between groups via common words much as in a sausage chain. More groups are recalled in such linked lists, and Ss clearly used the linkages in moving from one to another group in recall. A currently prevalent view of free recall learning is that it involves organizational processes (e.g., Bower, 1968; Mandler, 1967; TuNing, 1962, 1968). There are at least two aspects to this organization: first, the aggregation or grouping together of several words of the to-be-recalled list, so that these groups become functionally integrated units in recall; and second, the development by S of a retrieval plan or cuing scheme to enable him to move in recall from one functional unit to another in his memory. The following experiments attempt to manipulate the first factor, the subjective units or groups which S uses in free recall. There are several ways to study the role of subjective groupings in free recall. If a cluster of items is acting as a functional unit, then recall of the words in that cluster should exhibit a characteristic pattern; they should tend to appear close together in recall, perhaps even in a stereotyped order, and the interresponse times between recalling words in the unit will be short. The latency characteristics i The research reported in this article was supported by a grant, MH-13950, to the senior author from the National Institute of Mental Health. of this picture have been investigated by Pollio (1968), McLean and Gregg (1967), and Mandler (personal communication, 1968), who have found the recall pattern to be a rapid burst of responses in a subjective cluster, then a pause, another burst of responses from another cluster, another pause, and so on. The output stereotypy of words in subjective groups has been measured by Tulving's (1962) index of "subjective organization," or Bousfield, Puff, and Cowan's (1964) index of intertrial repetition. Both indices essentially measure the frequency with which S recalls pairs of words in a consistent order even though he is free to do otherwise. Although these indices are mathematically independent of the number of words recalled, both studies cited reported that these indices increased with practice on a list and correlated highly with the number of words recalled by individual Ss. Other methods for investigating these subjective groups are to have S write the w~rds as they are presented, allowing him to group them on his study sheet in any way he wishes (Seibel, 1967), or to have S sort the words, categorizing them into as many groups as he wishes before he recalls them (Mandler, 1967). 481

2 482 BOWER, LESGOLD, AND TIEMAN The conclusions suggested by these several methods are that Ss search for stable groupings of the list words, and that they tend to recall together those items which they have assigned to the same group. Furthermore, as multitrial learning proceeds, the assignment of items to groups becomes more consistent and the groups become better integrated in that more words can be recalled from each group. Tulving and Mandler have hypothesized that it is this increase in size and integration of the subjective groups that is the main cause of the increase over trials in the number of words which S can recall. If one believes that this is the cause for improvement in free recall with repeated trials, it then follows that if we somehow force S to change his groupings anew every trial, we should seriously degrade the usual increments in recall provided by repeated trials. This implication is tested in Exp. I. EXPERIMENT I In order to test this implication, one ideally would like to have some means for controlling how S groups a given set of words and, in particular, be able to force S to use different groupings on different input trials. Although it might be possible to construct lists of related words that we could get S to classify exhaustively in alternative ways on different trials, that method of testing the implication struck us as likely to produce results having only limited generality. We therefore developed a technique for imposing any arbitrary groupings which E wishes upon any arbitrary set of unrelated words (concrete nouns in this instance). The technique is elementary: S is merely shown a small group (2-4) of the words at one time and is told to make up an elaborate "mental image or picture" in which all the objects named are interacting together in some vivid and memorable scene. For example, with the quartet, dog, cigar, bicycle, hat, S might mentally imagine a scene in which a dog wearing a hat and smoking a cigar is pedaling a bicycle. From pilot work, we know that this procedure establishes very strong interassociations among these words, so they act much like an integrated unit. We also found that when we did this for quartets of words in a free recall list, we had practically complete control of S's clusters in free recall; although free to recall words in any order, S's entire protocol could be described as short runs of two to four responses from the quartet groupings presented to him. The cluster scores produced by this method (with unrelated nouns) are in fact much higher than the clustering scores one obtains with a highly categorized list of words presented one at a time in blocked fashion (e.g., Cofer, 1965). The clustering may be indexed by the modified ratio of repetition (MRR) which is ~njj MRR - ~(nj J - D' for all nj > 1, J where nj is the number of words recalled from category or group j, nj~ is the number of pairs of the nj words which are recalled in consecutive order, and the summation ranges over the various categories or groups. This clustering index is a fraction, anchored at 1 when there is perfect clustering. With the mental imagery method, we consistently obtain mean MRR scores of.90 or higher, representing almost perfect clustering in recall according to the groups E imposed upon S. It may also be reported that the latencies of successive responses in recall appear to follow the pauseburst-pause pattern reviewed earlier. However, we have not made extensive enough measurements of this to report any statistical data on latencies. With this method for controlling S's functional groupings, we are in a position to test the implication that multitrial free recall is retarded if the groupings of the words are changed on every trial. To change groupings, E merely rescrambles the list words and presents new quartets on each trial for S to associate together.

3 GROUPING OPERATIONS IN FREE RECALL 483 Method Ioo. Each S had three trials on each of four lists of 24 c~ tad nouns for free recall. For two of the lists, the quartet 2 90 groups were preserved intact over all three trials (Same t,.) " condition); for two of the lists, the groupings of words into quartets were systematically different on all three to 80 a trials with that list (Changed condition). The 24 words in each list were concrete nouns of high imagery ratings (Paivio, Yuille, & Madigan, 1968) chosen so as to be ~_ z 70 as Unrelated as possible. Across Ss, each list appeared "' CD rw equally often first, second, third, or fourth, and was equally often in the Same or Changed condition. ~- 60 l The 24 words of a list were divided randomly into 6 quartets for Trial 1. These 6 quartets were presented by a slide projector for 12 see. each (3 see. per word) with S instructed to mentally image an interactive scene for the four objects named. He was not told that he had to recall these four words together. After the sixth input slide, S began oral free recall for 60 see., being told to recall all words he could in any order he wished. Due to a procedural oversight, recall order was not recorded in this study. For lists in the Same condition, the same six slides of Trial 1 were presented again in the same order on Trials 2 and 3. For lists in the Changed condition, six new quartets were composed of the 24 words for slide presentation on Trial 2, insuring that no two words appeared in the same groups on Trials 1 and 2. Similarly, six further quartets were composed for presentation on Trial 3, insuring that no two words appeared together which had been together on Trials 1 or 2. The S was instructed to construct a composite mental image for whatever quartet of words was shown to him. A 30-second break plus instructional review were interspersed between the last trial of one list and the first trial of the next list. The Ss were 12 Stanford undergraduates fulfilling a service requirement for their introductory psychology course. They were run it~dividually. Results Data were pooled over word lists and over lists-within-session since neither of these incidental variables produced significant effects. The main results are shown in Figure 1 giving words recalled over the three trials for the Same vs. Changed conditions. Recall in these conditions began at the same level on Trial 1, as it should since the treatment was identical at that point. However, the conditions differed radically in their improvement over Trials 2 and 3 (p <.001). The list repeated with the same groupings improved markedly I 2 3 TRIALS PLES LES Fie. 1. Mean recall curves for Exp. I. with practice, whereas the list repeated with changing groups showed relatively little improvement. In fact, the null hypothesis of no improvement in recall in the Changed condition could not be rejected, F(2, 69)= 1.91, p >.20. The differences among conditions appear both in intertrial retention and intratrial forgetting (Tulving, 1964). Given that a word failed to be recalled on Trial n, its conditional probability of being recalled on Trial n + 1 was.87 in the Same condition and.61 in the Changed condition. Given that a word was recalled on Trial n, its conditional probability of being forgotten on Trial n + 1 was.04 in the Same condition and.22 in the Changed condition. Thus, the Changed condition was poorer in that fewer new words were picked up each trial, and more previously recalled words were forgotten each trial. Recall according to the clusters presented also differed considerably between the two groups. Figure 2 shows the probability histograms of words recalled per presented quartet, pooling the two conditions on Trial 1, and pooling Trials 2 and 3 separately for the two conditions. The Trial-1 histogram shows the "some-or-none" result reported by Cohen (1966) for free recall of categorized word lists, viz., S either does or does not get into a group or category (the high 0 point), but if he gets into it he recalls a high percentage of the

4 484 BOWER, LESGOLD, AND TIEMAN 1.0.6,& o 4- Pooled Trial 1 Conditions Words per Group Trials 2 and 3 Some Groups Changing / Group ~'/ / I T T,, 0 I Words per Group FIG. 2. Probability distributions of the number of words recalled per presented quartet, grouped over Subjects. words from that group. The distribution for the Same condition on Trials 2 and 3 shifts strongly towards "all-or-none" recall of the quartet, with predominately all of it being recalled. On the other hand, the distribution on Trials 2 and 3 for the Changed lists (scored relative to the groups presented on those trials) shifts away from the "some-or-none" pattern towards a simple linear distribution with the zero point falling in line along a gradient. This presumably means that the new quartets presented on Trial 3, for instance, are not acting as new functional units, but rather are conflicting with the prior groupings given on Trials 1 and 2. These results support the Tulving-Mandler hypothesis which led to this experiment. Improvement in multitrial free recall appears to be a concomitant of developing stable, integrated groups of list words; and if measures are taken to diminish stable groupings, multitrial free recall is seriously retarded. It is probable that Ss experiencing changed input groupings would eventually improve so as to recall nearly all the words. After a while, S may ignore the input groupings and establish his own output groupings, just as he apparently does in the usual free recall procedure. Future experiments shall have to determine the persistence with which changing input groupings can override S's demand for order in his output. The results contradict the view that free recall of a given word depends upon the number of other list words which are associated with it. By Trial 3 in the Changed condition, at least nine other list words had been associated with a given list word. Yet its recall was considerably poorer than in the Same condition in which only three other list words had been associated to a given word. However, association theory suggests an alternate explanation of our results; namely, that recall of a given word depends upon the number and strength of associations to it from other list words (e.g., Deese, 1959). On this view, items repeated in the same quartets develop strong associations to the three other items in that group, whereas items presented in new quartets develop weak associations to nine other items. Moreover, successive sets of three of these latter associations are related to one another by a general A-B, A-C (or A-Br) paradigm of negative transfer, perhaps causing unlearning of prior associations to an item. If the strength of associations to a given list word is a more important determinant of its recall than is the number of words associated with it, then this negative-transfer hypothesis might have some hope of accounting for the results. However, the specific details of an association theory for free recall are not yet worked out sufficiently to know whether it would be consistent with the present results. EXPERIMENT II In Exp. I, the input clusters were arbitrarily specified by E, and S apparently adopted these as functional units for recall. It is presumed, however, that if E did not group the items (i.e., presented them singly), then S would still come eventually to group them together in his recall, probably in an idiosyncratic manner. Suppose that at some point late in S's freerecall learning, we ask him to indicate the natural word groupings he has been using for his recall. With this knowledge of S's natural groupings, we would then be in a position

5 GROUPING OPERATIONS IN FREE RECALL 485 to predict whether a subsequent grouped input trial would increase or decrease his next recall depending upon the groupings imposed. If the clusters on this input trial are composed so as to correspond with the natural groupings which S has told us he uses, then his next recall should be facilitated by that method of presentation. However, if the input clusters are composed so as to systematically violate S's natural groupings of the words, then his next recall should be poorer than it was before, despite the additional input trial. The reasoning is much the same as for the Samevs. Changed-groupings comparison in Exp. I, except in the present experiment S's natural groupings of the words (rather than E's arbitrarily imposed groupings) serve as the reference point for presenting the Samevs. Changed-groupings of the words on the next trial. These predictions are tested in Exp. II. Method Each of 8 Ss from the previous source did two free recall and sorting tasks. On each list of 36 unrelated concrete nouns, S first had ungrouped input-output trials (at a 2.5-second input rate) until he reached a criterion of recalling at least 32 out of the 36 words (89 ~). After his criterion trial, S was then asked to sort the cards containing the 36 list words into 9 groups of 4 words each, putting together "those words which you have been recalling together or which you think of as belonging together in a group." The S was permitted as much time as he needed to do this, with all taking between 1.5 and 4 min. After recording S's groupings, E then gave one further input trial, this time with nine groups of four words (cards) presented together on the table top for 10 see. per quartet with mental imagery instructions given to S. For one of the lists (Consistent list), these 9 input quartets were exactly the same as the nine groups of four words that S had just indicated in his sorting. For the other list (Inconsistent), the nine input quartets systematically violated S's sorting categories, with each word in the quartet coming from a different one of S's sorting categories. After this input trial, S gave a final recall of the list. Each S received one Consistent and one Inconsistent list, half the Ss in that order, half in the reverse order. Over Ss, the two sets of 36 words served equally often as Consistent and Inconsistent lists, and equally often as first or second list being learned. Results Mean trial of reaching criterion was 4.12 for the Consistent list and 4.00 for the Inconsistent list; this is an insignificant difference. Mean words recalled on the criterion trial were 32.5 out of 36. Mean recall after the sorting and grouped input trial increased to 34.1 with the Consistent input trial and decreased to 30.2 with the Inconsistent input trial. The effect is small but it is quite consistent and reliable over the 8 Ss. All 8 Ss increased their recall after the Consistent trial (t = 5.65, p <.01, whereas 6 of 8 decreased recall after the Inconsistent trial (one increased by one word and one was unchanged), t = 2.57,p <.05. For all 8 Ss, recall was higher after their Consistent input trial than after their Inconsistent input trial, t = 4.78, p <.01. The effect of Consistent vs. Inconsistent groupings was relatively small here, but various considerations would have led one to expect this. First, the criterion performance of 32.5 out of 36 words is already very close to the true asymptote for this task, so one cannot reasonably expect much of an increment for the Consistent list. Second, the one Inconsistent input trial was working against the prior organizations established by an average of four preceding trials on the list plus the additional study time (mean was 2.5 rain.) provided by the lengthy sorting trial. The latter by itself would normally have been expected to produce an increase in recall but this apparently was overriden by the intervention of the inconsistently grouped input trial. The conclusion from this study is similar to that from Exp. I: S's recall is poorer if the input groupings of the words are different from the familiar groupings he has previously been using for guiding his recall. The difference between Exp. I and the current one is that in the former case the prior word groupings were imposed by E whereas in the latter case these presumably were developed by S and were indicated by his sorting behavior.

6 486 BOWER, LESGOLD, AND TIEMAN EXPERIMENT III The third experiment inquired whether changing word groupings over trials would affect recognition memory or discrimination of list membership. Integration of several words into a higher-order unit presumably affects the probability of retrieval, since any item of the cluster may cue recall of any other; but conceivably, recognition depends only upon some "occurrence" information stored with each item that is independent of interitem associations. For example, Kintsch (1968) has shown that blocked, categorized lists are better recalled than unrelated word lists but that they do not differ in recognition. In his terms, categories, or the interitem associations among items in the categories, affect retrieval processes, but do not affect the basic discrimination that a word was on the list. Perhaps we shall be able to show a similar separation of recognition and recall for our lists that are grouped in the same way or in changing ways over trials. Method The basic procedure was to give S two input-output cycles of free recall on a list in which the groupings were the same or different on the two input trials, and then have a recognition test after the second output trial. Each S learned two lists in a counterbalanced order; for one list, the groups were the same on Trials I and 2; for the other list, the groupings changed. To avoid a ceiling on recognition performance, the lists were lengthened to 75 words, and the presentation time was 3 sec. per word. Triplets of words were presented for 9 sec. with mental imagery instructions. There were 25 such triplet slides in the list. For Same-groupings lists, the same 25 triplets were repeated on the two input trials; for Changed-groupings lists, the 75 words were arranged into 25 new triplets for presentation on Trial 2. Recall was in writing, with 3 rain. allowed. Immediately after the second recall of a given list, the recognition test was given. The S was given a sheet of paper listing in random order the 75 list words mixed in with 75 synonymic or closely related distraetors. He was told: "Some of these words were on the list you have just studied and some were not. Check off those words which you think were on the list." The Swas not told how many list words were on the sheet nor how many words to check off. Each list word had a corresponding synonym or related distractor. In order to achieve 150 related pairs (75 on two lists), a few abstract nouns were used as well as concrete nouns. The Ss were 16 students from the previous source. They were run individually with counterbalancing over Ss of which word list was first or second and whether a Same or Change list was first or second. Results The results are summarized in Table 1 pooled over the variables of word lists and first vs. second list in the session. Perhaps due to the greater number of functional groups (25 vs. 6) and the presence of some abstract words, recall in Exp. IV was considerably less than in Exp. I. The two condkions, Same vs. TABLE 1 FREE RECALL AND RECOGNITION SCORES FOR SAME VS. CHANGED CONDITIONS Same groupings Changed groupings Free recall Trial Trial Increment Recognition Hits False alarms Changed lists, were similar in recall on Trial 1 but differed on Trial 2, t(15) = 2.82, p <.01. The increment in recall from Trial 1 to Trial 2 was significant for both conditions, but the increment was larger for the Same condition (.28) than for the Changed condition (.14), t(15) = 3.60, p <.001. Thus, the recall portion of this experiment replicates the qualitative finding of Exp. I, that the trial-to-trial improvement in free recall is less with changed groupings of the words. In contrast to the recall differences, there were no differences in recognition between the Same vs. Changed conditions. Hit rates (checks on list words) did not differ, false alarm rates (checks on nonlist words) did not differ, nor did the hit-minus-false alarm scores (latter t(15) = 1.09, p >.15)~ The conditional relationships between recall

7 GROUPING OPERATIONS IN FREE RECALL 487 on Trial 2 and recognition of a word were examined for the two conditions. If a word was recalled on Trial 2, its probability of being recognized was.93 in the Same condition and.97 in the Changed condition. If a word was not recalled, its probability of being recognized was.65 in the Same condition and.61 in the Changed condition. The difference in recognition of previously recalled words in favor of the Changed list was sufficient to offset the difference in recall, thus yielding near equivalence in net recognition despite the differences in recall of the two lists. The net outcome of this experiment is exactly the pattern conjectured on the basis of Kintsch's (1968) prior results. Variables affecting stability of groupings of list words influence recall but not recognition. The account of these results is similar to Kintsch's. When a word is studied, S stores information about its occurrence on the list (perhaps as a recency or contextual frequency tag or as a trace strength), and he also tries to learn some method for retrieving that item in recall (perhaps associating it to a category cue, or to other list words). Grouping and regrouping operations affect the efficacy of the latter sorts of retrieval information, but have no effect upon the occurrence information stored alongside the word in S's lexicon. And it is the latter sort of information that is consulted when S is shown a word and asked to decide whether it was on the list just studied. EXPERIMENT IV The organizational hypothesis of Tulving and Mandler implies that multitrial free recall increases in part because the subjective groups increase in size over trials; the clusters presumably become larger and better integrated with practice. In the following experiment, by use of varied input groupings, we have tried to facilitate or retard this growth over trials in the size of subjective clusters. In the Increasing condition, the input groupings increased in size over the three trials, with individual groups on Trials 2 and 3 composed of two intact groups from the preceding input trial. In the Decreasing condition, the group sizes correspondingly decreased over the three trials, by halving on Trials 2 and 3 the groups of the preceding input trial. For the Increasing condition, the group sizes were over the three trials; for the Decreasing condition, they were The hypothesis expects that Ss having increasing group-sizes will recall better than Ss having decreasing group sizes. Method Twenty paid Ss (Stanford students) had three inputoutput cycles on a list of 60 concrete nouns. These words were randomly composed into 20 triplets and typed on 4-6-inch cards; then two random triplets were combined to make a list of ten 6-tuples; then two random 6-tuples were combined to make a third list of five 12-tuples. By appropriate spacing, the 6-tuple and 12-tuple cards were clearly divided into the two or four constituent triplets of words. The cards were presented manually to S. Ten of the Ss received these list groupings in the order on input Trials 1, 2, 3 (Increasing condition), and ten Ss received the reverse order of groupings over trials (Decreasing condition). The Ss were instructed to try to form an elaborate mental image or imaginary scene in which au the objects named on a given card were interacting together in some vivid or memorable way. Presentation time was calculated at 4 sec. per word, so 3-tuple cards were shown for 12 sec. each, 6-tuples for 24 sec., and 12-tuples for 48 sec. After presentation of the last input card, S began free recall orally for 2 min., being instructed to recall as many words as he could in any order he preferred. Results The proportions of words recalled over trials by the two groups are shown in Figure 3. The groups are equal in recall on Trial 1, but the Increasing Ss improve at a faster rate than do the Decreasing Ss. Although the difference in total recall is not large in an absolute sense, it is highly significant statistically, F(1, 54)= 10.96, p <.005. We next examined these data for group differences in clustering. There are a variety of ways to look at this feature of the data. One

8 488 BOWER, LESGOLD, AND TIEMAN ~o o 6O "E " 3C In ore a sin~.//////.~.,, Trials FIG. 3. Mean recall curves for the increasing (3-6-12) and decreasing (12-6-3) conditions of Exp. IV. statistic is the MRR index presented earlier. These MRR scores calculated for 3-tuples, 6-tuples and 12-tuples were consistently higher for Increasing Ss than for Decreasing Ss, indicating greater conformity to the input clusters by the Increasing Ss. A second statistic we have examined is the conditional probability that S recalled all the words of a group given that he recalled at least one word of that group. This is a convenient index of the integration of the group. Since the definition of an input "group" changes systematically over trials for each condition, we have scored recall for each definition of "group" on each trial. For example, on Trial l for Increasing Ss who had input triplets, we scored for triplets and also for the 6-tuples and 12-tuples they were ~ 01 Increasing Groups LO I Decreasing Groups ~- O''r.. I 2 3 I 2.'5 Trials Trials FIG. 4. Conditional probabilities of subjects recalling all of a group given that they recalled part of that group (Exp. IV). going to see on later trials. This index of cluster integration is depicted in Figure 4 for the three cluster sizes, for the Increasing condition on the left and Decreasing condition on the right. Within each panel, the integration index is necessarily ordered from 3-, to 6-, to 12-tuples. The salient conclusion from comparing the two panels of Figure 4 is that group integration is markedly higher for the Increasing Ss than for the Decreasing Ss on all trials for all sizes of groups. Could these clustering results be an artifact simply of differences in mean recall? The percentages can be "corrected" on the assumption that recall of each word in the list occurs randomly with probability p, on Trial n (cf. Figure 3). If individual words were recalled at random, then the probability of recalling all of a group of size k given recall of at least one word of that group would be p,k/[1 - (1 _p,)k]. After making this correction for the mean recall level, two conclusions remain: first, that clustering according to input groupings is still very much greater than one would expect from random output at a given average probability level; and second, that the clustering in excess of chance is still much greater for the Increasing Ss than for the Decreasing Ss. Although the results of Exp. IV came out significantly in the expected direction (Increasing better than Decreasing), it must be admitted that the effect is not large in an absolute sense. Why might this be? We think the answer to this must refer to certain organizing activities of S that simply are not controlled by our input-grouping procedures. For example, although presentation of triplets controls fairly welt the tendency for S to recall these words as a unit, it does not control the order in which he recalls the triplets nor does it prevent his organizing (grouping, associating) several triplets into a larger subjective unit for his recall. Consequently, when for the Increasing Ss we aggregate two triplets at random for a 6-tuple presentation on Trial 2, there is a good like-

9 GROUPING OPERATIONS IN FREE RECALL 489 lihood that we will aggregate triplets differently from how S did in his Trial-1 recall. Similarly, in the Decreasing condition, when we divide a 6-tuple from Trial 2 into two 3-tuples for Trial-3 presentation, we have done nothing to prevent S from continuing to use his old association between these 3-tuples in recalling on Trial 3. The point of these remarks is to suggest that the groupings we impose on a particular input trial cannot be the only organization that S will have available in his recall on that trial. To the extent that S rather than E controls some of this organization, manipulation of Increasing vs. Decreasing group sizes will have an attenuated effect on recall. These considerations may help explain some of the puzzle of why these groupsize manipulations produced only a small difference in recall. EXPERIMENT V If S forms subjective units or clusters of words~ he still has the problem of getting from one cluster to the next in his free recall. One way this can be achieved, it is presumed, is by associating one cluster with another, or by aggregating them into a larger cluster as in Exp. IV. In the following experiment we have tried another method to aid and abet the S in moving from one cluster to the next in his memory. We tried to do this by arranging for structural linkages between two clusters via a common element. This common element could then serve as a mediating bridge for moving in memory from one cluster to the next during recall. The structures of the Linked list and Control list are schematized in Figure 5, where letters represent unrelated nouns of the list to be recalled and ellipses are drawn around the input quartets. The quartets of the Linked list are linked together by common elements (A, D, G, J) which appear in two different clusters. All quartets were linked at both ends in this way, much in the manner of a sausage chain with its end attached back upon its LINKED LIST CONTROL LIST (A D 6 J) (B C E F) (H'KO QA D G J) FIG. 5. Structure of presentation lists for Exp. V. beginning. Although Figure 5 shows only 4 "sausage links," there were 16 in the actual experiment. They were presented in a random temporal input order although they were linked structurally in a continuous chain. Examining the Control list in Figure 5, it shares with the Linked list the fact that the same words (A, D, G, J) are presented twice and the same words once. Its difference from the Linked list is that no common linking elements are provided to enable S to move in memory from any one cluster to any other, and S is left to do this by his own devices. The prediction is that Ss learning the Linked list will recall more clusters, and therefore more words, than Ss learning the Control list; moreover, the output order of the linked clusters should correspond in large degree to the succession of clusters as linked in the sausage chain. A preliminary test of this with a list containing eight quartets revealed no differences in recall between a Linked and Control list. This worried us until we scored the protocols for cluster recall, and discovered practically all Ss in both conditions were recalling some words from practically all of the eight input clusters on Trial 1. However, the hypothesis supposes that any advantage that might appear for the Linked list would have to be in terms of more clusters recalled than in the Control list. But since all eight clusters were being recalled on the Control list by practically all pilot Ss, there was no possibility of showing an advantage for the Linked list in that situation (i.e., a ceiling effect). To avoid an artificial ceiling, therefore,

10 490 BOWER, LESGOLD, AND TIEMAN we ran the main experiment reported below with a longer list of 16 clusters. Method Twelve Ss learned two Linked lists and 12 other Ss learned two comparable Control lists. Each list was given for three input-output cycles. Each list contained 16 concrete nouns presented twice, and 32 concrete nouns presented once. The words, grouped into 16 quartets, were presented by a slide projector, at a rate of 12 sec. per slide with visual imagery instructions to S. The 16 slides for a given list were shown in a different random order over the three trials. The structure of the Linked and Control lists was as illustrated in Figure 5. After the last study slide, S wrote his free recall for 4 min. There was a rest pause of 1 min. between the end of S's first list and the beginning of his second. The Ss were run individually and. were not informed of the structural arrangements of the lists they were to learn. Over Ss, the two lists of 48 words were used equally often as first or second lists in the session. Results There were only small differences between recall of the two lists of the session, so they have been pooled to increase reliability. The main results are shown in Table 2, giving mean TABLE 2 MEAN RECALL PROBABIL~S : EXPERIMENT V Single Double Trials Linked Control Linked Control recall probability on each trial for the two conditions for the once-presented (singles) and for the twice-presented (doubles) words. The doubles are recalled much better than the singles, of course, as any theory would expect. Although most of the recall differences between the Linked and Control lists are in the predicted direction on each trial, the only significant difference is on Trial-1 recall of the once-presented words, t(22) = 2.63, p <.01. The total recalls of the once-presented words over all three trials is also significantly higher in the Linked list, t(22) = 1.82, p <.05. We next analyzed for recall of the input clusters. In recall of the Control list, the average MRR scores was.987 on Trial 1,.990 on Trial 2, and.993 on Trial 3, reflecting almost perfect output clustering according to the input groups. For the Linked list, MRR is not easily computed because of the double words which appeared in two different clusters. The hypothesis implies that the advantage for the Linked list is in S's ability to recall some words from more clusters than is true for the Control list. One therefore wants to compare the number of clusters recalled (i.e., at least one word of the cluster) for the two lists. However, because doable words in the Linked list appear in two different input groups, if S recalls, say, DBC (see Figure 5) from the Linked list, we do not know whether to credit him with two clusters (since D appeared in two) or one cluster recalled. To avoid all problems of this kind, we decided to redefine a recall cluster in terms of the two unique words in the input quarters; examples in Figure 5 are BC, EF, HI, and KL. If S recalled either member of these 16 unique pairs, he was credited with recall of that cluster. Recall of the Control list was scored in the same manner so that, for instance, the unique quartet BCEF in the Control list of Figure 5 was redefined as consisting of the two clusters BC and EF. Regarding the prediction of interest, this method of scoring is conservative, and gives an advantage to the Control group, since recall of one of their input groups is being artificially scored as recall of two redefined clusters. Nonetheless, the differences between the two conditions came out quite substantially in the predicted direction. Proportion of these clusters recalled on Trial 1 was.53 for the Linked list vs..32 for the Control list, p <.01 ; on Trial 2, the proportions were.82 vs..74, p <.05; on Trial 3, they were.86 vs..88. Thtls, even with this conservative scoring method, duster recall was higher on early

11 GROUPING OPERATIONS IN FREE RECALL 491 trials for the list which provided explicit linkages between input clusters. The second prediction is that S's temporal order of recalling the linked clusters will correspond to a significant degree with the succession of "sausage links" in the structural chain, unwinding it in either a clockwise or counterclockwise direction. By casual inspection of the recall protocols, this is indeed an accurate description, but to develop a quantitative index of this ordering is a ticklish problem. After considering various ways to measure this, we finally hit upon the following method which proved sufficient to show what was obvious from inspection of the recall pro~:ocols. The unique words were simply ignored in recall of the Linked list, and only recalls of the double words were considered. These were numbered from 1 to 16 according to their structural succession in the input list. For example, in Figure 5 the double words A, D, G, J would be numbered 1, 2, 3, 4, respectively. By this coding, a S's protocol would consist of a sequence of integers, such as 7, 8, 1, 16, 15, 9, 10, 4, 3, 5. We wished to measure the pairwise correspondence between such recall sequences and what one would obtain if one cycled through the numberloo p 1-16 either forwards or backwards. Consecutive pairs of numbers in the recall sequence were therefore scored as plus or minus according to whether or not that pair of numbers were adjacent to one another in the natural numerical order (considering 1 and 16 as adjacent). For example, the sequence above has 9 consecutive pairs of which 5 are plus and 4 are minus. The number of pluses will be called the "succession score." All 72 Linked-list recall protocols (12 Ss x 2 lists x 3 trials) were scored in this manner. The number of pluses varies according to the. number of double words recalled and which ones they are. The question is whether the obtained succession scores exceed those to be expected by chance if the sequential ordering of the double words recalled were totally random. We answered this question with a computer by running 500 random Monte Carlo permutations of the recall number for each of the 72 protocols, counting the successions in each random permutation, and thereby developing a probability distribution of the succession scores to be expected by chance if the words recalled in that protocol had been generated in a random sequence. In this way, we were able to ascribe to each of the 72 protocols (double-word sequences) the probability that a succession score as high or higher than that observed would have occurred had the order of recalled words been governed by a totally random process. These results can be summarized in terms of the number of protocols for which the observed succession score had, on the null hypothesis, a theoretical probability less than.01, or.05, or.50. Thirty-four protocols, or 47 ~, had p values less than.01; 50, or 69 ~, had p values less than.05; and 65, or 90~, had p values less than.50. Quite obviously, these outcomes mean that the recall protocols were strongly ordered in proper succession. Subjects were indeed recalling around the "sausage links." In conclusion, by arranging linkages between input clusters, Ss recall more clusters on the early trials, and they clearly output the clusters in an order prescribed by the structural array. This occurred despite the fact that these clusters were shown in a random temporal order. This benefit for recall has been shown where the linkages were established by a common word in two clusters. One probably would be able to demonstrate a similar effect with highly associated pairs of words, with the two members of associated pairs being embedded in different input clusters. Thus, recall of black in its cluster might cue recall of white and its input cluster. That experiment is yet to be done. The advantage demonstrated here for the Linked list was small and short lived, disappearing by Trial 3, presumably because Control Ss were recalling from nearly all of their input clusters by that trial. Possibly the

12 492 BOWER, LESGOLD, AND TIEMAN advantage would prove more persistent with a longer list of clusters. The size of the Linkedlist advantage may have been attenuated to some degree because the linking words were each presented for association with two different sets of words on each trial. This within-trial procedure for the linking words therefore shares some of the features of the between-trial changes in groupings which produced, deleterious effects on recall in Exp. I and IV. However, the shifting groupings in Exp. I and IV were much more extensive than here, and the present experiment repeated the identical groupings over all three trials. DISCUSSION To recapitulate the argument and our experimental data: we believe that a fundamental strategy Ss employ in learning is to group or subdivide the material into subjective clusters which become integrated units in recall. These strivings for stable groupings can be assessed in various ways which tap the organization S employs in free recall. The hypothesis is that the improvement with practice in free recall results in part from the increasing size and integration of these subjective clusters. It follows that recall should be retarded if measures are taken to prevent the development of stable groupings with practice. Using the "mental imagery" method for interassociating groups of four unrelated nouns, Exp. I showed that the normal improvement with practice in free recall practically vanished when new groupings were imposed upon the list words on each trial. In Exp. II, rather than E imposing arbitrary groupings, S indicated his subjective groupings of the words developed during preliminary recall trials. A subsequent input trial with imposed groupings increased or decreased recall accordingly as the imposed groupings were consistent or inconsistent with S's subjective groupings. We view these two experiments as establishing the same point: If current input groupings conflict with prior groupings of the material, then recall suffers. The experiments differ only in the nature of the prior groupings and how they were established. The third experiment replicated the poorer recall learning with changed groupings but showed that the effect was absent in recognition tests of memory. It was proposed that grouping factors influence response generation and retrieval processes, but not recognition since the latter depends only upon "occurrence information" stored in memory for each word, which information is independent of retrieval cues or schemes for accessing that word in recall. This "occurrence information" is presumed to be accessed directly by the word on a recognition test. Given that S has categorized the list words into many subjective units, he still has the problem of getting to all these groups in recall. Mnemonic techniques use various (cuing) methods for solving this retrieval problem. A less dramatic but all-purpose method is to associate two or more groups together, integrating two former units into one larger chunk, thus reducing the number of "units" to be retrieved. In Exp. IV, this composition of subgroups into larger groups was either aided or hindered somewhat by the groupings presented over successive trials, and recall was better in the former case, as expected. In Exp. V, S was aided in moving in memory from one to another recall cluster by the presence of a common word linking "adjacent" clusters. These structural linkages improved cluster recall; it was further demonstrated that S clearly used these linkages in moving from one to another cluster in his recall. Our initial assumption was that free recall reflects in part the amount of grouping which S has carried out on the list words. This view is bolstered by several incidental learning experiments with free recall. Tulving (1966) found little effect on tree recall of having S merely read the list words many times before being told they were to recall the words.

13 GROUPING OPERATIONS IN FREE RECALL 493 Contrariwise, Mandler (1967) found that incidental Ss, required to group the list words into consistent categories, recalled as much as intentional Ss told to categorize then recall. Considered more generally, however, it is not clear what kinds of categorization will or will not benefit recall. For example, if words are classified by their number of letters, by whether or not they contain a t, by their part of speech, etc., one intuitively would expect little incidental recall from such activities. The important processes probably involve the arousal of semantic features of the word, and the arousal of categorizations that have strong associations to the list words. REFERENCES BOUSFXELD, W. A., PUre, C. R., & COWAN, T. M. The development of constancies in sequential organization during repeated free recall. Journal of Verbal Learning and Verbal Behavior, 1964, 3, BOWER, G. H. Organization and memory. Address at meetings of Western Psychological Association, San Diego, California, March, COFER, C. N. On some factors in the organizational characteristics of free recall. American Psychologist, 1965, 20, COHEN, B. H. Some-or-none characteristics of coding. Journal of Verbal Learning and Verbal Behavior, 1966, 5, DEESE, J. Influence of inter-item associative strength upon immediate free recall. Psychologica! Reports, 1959, 5, Kiyrscn, W. Recognition and free recall of organized lists. Journal of Experimental Psychology, 1968, 78, MANOLER, G. Organization and memory, in K. W. Spence & J. T. Spence (Eds.), The Psychology of Learning and Motivation, Vol. 1. New York: Academic Press, McLEAN, R. S., & GREGG, L. W. Effects of induced chunking on temporal aspects of serial recitation. Journal of Experimental Psychology, 1967, 74, PAIVIO, A., YUILLE, J. C., & MADIGAN, S. A. Concreteness, imagery, and meaningfulness values for 925 nouns. Journal of Experimental Psychology Monograph Supplement, 1968, 76, No. 1, Part 2, POLLIO, H. R. Associative structure and verbal behavior. In T. R. Dixon and D. L. Horton (Eds.), Verbal Behavior and General Behavior Theory. Englewood Cliffs, N.J. : Prentice-Hall, Inc., Pp SEmEL, R. Organization in learning. Tech. Rep., Contract No. OE Pennsylvania State University, University Park, Pa., TULXaNG, E. Subjective organization in free recall of "unrelated" words. Psychological Review, 1962, 69, TOLVING, E. Intratrial and intertrial retention: Notes towards a theory of free recall verbal learning. Psychological Review, 1964, 71, TULVl~G, E. Subjective organization and effects of repetition in multi-trial free-recall learning. Journal of Verbal Learning and Verbal Behavior, 1966, 5, TULVIN~, E. Theoretical issues in free recall. In T. R. Dixon and D. L. Horton (Eds.), Verbal Behavior and General Behavior Theory. Englewood Cliffs, N.J. : Prentice-Hall, Inc., Pp (Received December 16, 1968)