40 .311 * .115*. Lo-Lo Word Pairs dirt. WOOL. A .000* .000* smooth. FLOOR. AA ... MUSIC. AA .022*t .164. Hi-Lo Word Pairs ear. DOOR. AA .034 .000* eagle.
Memory & Cognition 1979, 7 (3), 224-231
Retrieval operations in cued recall and recognition RONALD P. FISHER Erindale College, University 0/ Toronto, Toronto, Ontario, Canada The present study compared the effects of two classes of experimental manipulations on the recognition and cued recall of target words learned in the presence of list cues. One dass of manipulations, learning instructions (repetition vs, meaningful rehearsal), had similar effects on recall and recognition, whereas the other, preexperimental association between target and cue words, bad separable effects: Cue-to-target association affected only recall, while target-to-cue association affected only recognition performance. Recall and recognition were thus viewed as fundamentally similar processes, both of which require retrieval operations. Differences between the two were attributed to the differential abilities of the recall and recognition retrieval cues to access the original episodic event. Most of the earlier attempts to explain recallreeognition differenees were founded on the assumption that reeognition is inherently easier. Consequently, these explanations were quantitative in nature. F or example, various "strength" theorists postulated that the decision proeess in the reeognition test is based on a lower threshold than in the reeall test (Bahrick, 1965; Postman, Jenkins, & Postman, 1948). More recently, different versions of the "two-process (generatereeognize)" theories suggested that reeognition is easier beeause an additional retrieval operation is required to sueeessfully reeall an event than to reeognize it (Anderson & Bower, 1972; Kintseh, 1970). The recent work of Tulving and his eolleagues (Tulving & Thomson, 1973; Watkins & Tulving, 1975), however, demonstrates that under a variety of eonditions, reeall is either no worse or is better than recognition. Differences between the two, then, eannot be thought of solely in quantitative terms, sinee reeognition is not inherently any easier than recall. A different approach to the problem is to eompare the two along a qualitative dimension. The most obvious plaee to start the search is with the different retrieval information used in the two paradigms: In reeall a contextual cue is provided, whereas in recognition a eopy of the target item is provided. If memory behavior is to be determined, at least in part, by the extent to whieh the information in the retrieval eue overlaps or aeeesses the information in the eneoded traee (Loekhart, Craik, & Jacoby , 1976; Tulving, 1974), then it is reasonable to expeet that under some eonditions, This research was supported by Grant A826I from the National Research Council of Canada to Fergus Craik. I would like to thank Fergus Craik and Eileen Simon for their helpful comments on an earlier draft. Requests for reprints should be sent to Ronald Fisher, Department of Psychology, Florida International University, Tamiami Campus, Miami, Florida 33199.
Copyright 1979 Psyehonomie Society , lne.
a eontextual eue may be more effieient than a eopy of the target. That is, there is no apriori reason why a eopy of the target item must be more sueeessful in aeeessing the information eontained in the eneoded traee than is a eontextual eue. An implicit assumption of this approach is that the eneoded traee is not simply a repliea of the nominal target, but that it is the result of whatever eognitive operations are performed on the target and its eontextual environment (Anderson & Bower, 1972; Loekhart et al., 1976; Tulving, 1976). The present study makes no attempt to eritieally test different meehanisms to aeeount for the effeets of informational overlap between the eneoded traee and the retrieval eue. lnstead, this prineiple serves as a perspeetive from which to view memory behavior, and it leads to the following argument. In the present context, there are three information sources that may overlap: the eneoded traee, the reeall eue, and the reeognition eue. Reeall performance is related to the overlap between the eneoded traee and the reeall eue, and reeognition is related to the overlap between the encoded traee and the reeognition cue. lf the inforrnation eontained in the reeall and reeognition eues ean be manipulated independently, then there should be some factors that affeet reeall but not recognition performance and others that affect recognition but not reeall. That is, any factor that affeets only the copy cue's ability to aeeess or overlap the eneoded traee will affect only reeognition performance. Conversely, any factor that affeets only the contextual eue's ability to aeeess or overlap the eneoded traee will affeet only reeall performance. For the purposes of translating these ideas into an experimental format, eonsider Bartlett's (1932) suggestion that one's reeolleetion of an event is the produet of his elaboration of the retrieval information. This elaboration or reconstruetion is guided by the preexperimental storehouse of information (semantic
224
0090-502X/79/030224-08$01.05/0
RECALL AND RECOGNITION memory) about the retrieval eue, in partieular, its relation to the desired episodie traee. The more aeeessible the preexperimental assoeiation between the retrieval eue and the eneoded episodic traee, the more accurate will be the reeonstruetion. In this respe et it is assumed that reeall and reeognition funetion in sirnilar fashion. Differenees between these two forms of memory arise from the different retrieval eues available: In reeall, the retrieval eue is a eontextual one, whereas in reeognition, the retrieval eue is a nominal eopy of the target item. It is expeeted, therefore, that the preexperimental association from the eontextual eue to the information in the episodie event (target plus eontext) should affeet the ability to reeall the original event, whereas the preexperimental association from the target item to the information in the episodic event should affeet the ability to reeognize the original event. These notions about the effeets of preexperimental associations between the information in the retrieval eue and the eneoded event were tested in the following manner. Subjeets were indueed to learn aseries of target words, eaeh in the eontext of another eue word, so that the eneoded traee was some eomposite of the target plus the eue words. These target-eue pairs were seleeted so that the preexperimental assoeiation from the eue to the target word was either high or low; that is, the eneoded event (target plus eue) was either easily generated from the eue word (high cue-to-target assoeiation) or diffieult to generate from the eue word (low eue-to-target assoeiation). It was expeeted that this
225
manipulation would affeet only eued reeall performance. A seeond set of target-plus-eue word pairs was seleeted so that the preexperimental assoeiation from the target to the eue word was either high or low; thus, the original event was either easy or diffieult to generate from the target word. It was expeeted that this manipulation would affeet only reeognition performance.
EXPERIMENT 1 Method
Stimuli. Each word pair consisted of a target word printed in uppercase letters and a cue word in lowercase letters. There were 48 critical word pairs that were equally divided into four types: Hi-Hi- Both the target and cue words mutually elicited each other with high probability in a preexperirnental free association test (e.g., window-GLASS); Lo-Lo-the target and cue words mutually elicited each other with low probability in the free association test (e.g., bath-KING); Lo-Hi-vthe cue word elicited the target word with low probability, but the reverse association occurred with high probability (e.g., catLION); Hi-Lo-vthe cue word elicited the target word with high probability , but tOO reverse association occurred with low probability (e.g., cabbage-HEAD). The preexperirnental free association data were taken from the Keppel and Strand (1970) and Palermo and Jenkins (1964; college group) norms, such that the criterion for high probability of elicitation was at least 3.3% of the responses; the criterion for low-probability associations was less than 1.1% of the responses. The median probability of elicitation was .092 for the highprobability associations and .000 for the low associations. All target words were high-frequency nouns (Thorndike-Lorge, 1944, count of at least 24/million) with target word frequency approximately equated across word types (see Table 1 for a listing of all the target-eue pairs),
Table I Target-Context Word Pairs Probability of Preexperirnentally Eliciting Context Word
Target Word
Target Word Frequeney
Target from Context Word
Context Word from Target
.042 .093* .088* .044* .042 .107 .253* .159 .093* .033* .216 .311 *
.075 .106 .047 .099* .038* .264* .143 .346* .264 .072 .099* .115*
.000* .001 .000 .000* .001 .000 .000 .001 .001 .000 .000* .000
.000* .000* .000 .001 .000 .000 .000· .000· .006· .000· .001 .000
Hi-Hi Word Pairs health train nail law sleep sour circle sweet steal snow window earpet
DOCTOR WHISTLE FINGER JUDGE NIGHT LEMON SQUARE CANDY THIEF MOUNTAIN GLASS RUG
dirt smooth bath hair stove command quiet trouble stand bed dry dream
WOOL FLOOR KING NEEDLE LAMP CITY SKY BOOK CROWD SHOE EARTH BUTTERFLY
AA
A AA
AA AA 27 AA
32 28 AA AA 40 Lo-Lo Word Pairs A AA AA 34 A AA AA AA AA AA AA 22
226
FISHER Table 1 Continued Probability of Preexperimentally Elieiting Context Word
Target Word
Target Word Frequeney
Target from Context Word
Context Word from Target
food eat blaek eat man gun hard soft yellow house smoke song
TABLE STOMACH SPIDER LION PRIEST SOLDIER HAMMER BUTTER CHEESE COTTAGE TOBACCO MUSIC
Lo-Hi Word Pairs AA 30 24 A 42 AA 34 AA 45 46 36 AA
.000* .000* .000* .000* .000 .008 .000 .001 .002 .002 .000* .022*t
.059 .092 .055 .050 .066 .091 .042 .054 .071 .264 .482 .164
ear eagle ehair mutton citizen street blossom baby stern blue moon eabbage
DOOR BIRD LEG LAMB PERSON LIGHT APPLE GIRL PLANT WATER STAR HEAD
Hi-Lo Word Pairs AA AA AA 45 AA AA A AA AA AA AA AA
.034 .576 .035 .317 .122 .057 .090 .045 .201 .034 .239 .071
.000* .006* .006* .000* .000* .011* .000* .000 .006* .006* .Oll .004
Note- Target word frequency = Thorndike-Lorge (1944) G count. -Elicitation probability is from Keppei and Strand (1970) norms (all others are from Palermo & Jenkins, 1964, norms). tExceeds criterion for low probability of elicitation (.011). Procedure, All of the subjeets were initially given a setindueing task in whieh 20 target-eue pairs were presented at a 5-see rate. The subjeets were instrueted to learn the words in upperease letters in terms of the eontext of the other word on the stimulus eard. After the last word pair, a eued reeall test was given in whieh the eue words were presented individually as the retrieval eues, and the subjeets were to write the associated target words. Subjeets were allowed 8 sec per retrieval eue to make the appropriate response. After the initial task, the subjeets were divided evenly into a reeall and a recognition group. Both groups were presented with 54 word pairs at a 5-see rate. Immediately after the last pair, the reeall group received a eued recall test in whieh the list eue words were presented individually as the retrieval cues; the subjeets were given 8 sec per retrieval eue to write the appropriate target word. Those in the reeognition group were given a test sheet containing the 48 target words that had been presented plus 144 foils. The subjeets were given unlimited time and were instrueted to cirele the 48 target words they had just learned; all subjeets were asked to make exaetly 48 responses. Shortly after this test, half of the subjeets in the recognition group were given a eued reeall test of the same word pairs; this test was identieal to that reeeived by the reeall group. Subjeets and Design. Forty-eight Erindale College students of both sexes served as paid volunteers: 24 were assigned to the reeall group and 24 to the reeognition group. Only the last 12 subjeets in the recognition group reeeived the additional eued recall test. The subjeets in both groups saw the identieallist of 54 word pairs on the data eolleetion trial. The middle 48 word pairs, whieh were later tested, were randomly ordered with the eonstraint that every block of 4 word pairs contain one member from eaeh word pair type: Hi-Hi, Lo-Lo, Lo-Hi, and Hi-La. Two
different random orders were used in the experiment. In the recall test, the order of testing matehed the input order so that a constant presentation-test lag obtained for all word pairs. In the recognition test, the 48 target words were randomly dispersed among the 144 foils. The foils and targets were matched for word frequency, so that for every target there were three foils of similar frequency.
Results For the recall tests, a response was scored as correct only if the target word was given as the response to the appropriate cue word. (Fe wer than 3% of all the responses were intralist intrusions.) In the recognition test, all the subjects complied with the instructions to make exactly 48 "old" responses. The recall (recall group) and recognition scores for the four word-pair types are presented in Table 2. As expected, recall performance was direct1y related to the preexperimental association from the cue word to the target [F(1,23) = 51.7, P < .01], whereas the reverse association, target to cue, had no effect [F(1 ,23) = 1.1, p > .25]. On the other hand, recognition scores were directly related to the association from the target to the cue word [F(1 ,23) = 53.0, p< .01], but there was no effect of the reverse association, cue to target [F(1 ,23) = 1.3, r > .25]. Both two-way interactions, Cue to Target by Test and Target to Cue by Test, were reliable [F(1 ,46) = 48.9, p < .01, and F(1 ,46) = 17.8, p < .01, respectively].
RECALL AND RECOGNITION
227
Sinee a different set of 12 target words was used for Table 2 Cued Recall and Recognition Accuracy of eaeh of the word-pair types, the possibility exists that Word Pairs in Experiment 1 the observed differenees are due only to the differential reeallability and reeognizability of the target words in Cue-to-Target Association the four word-pair types. For example , the Hi-Hi word Target-toRecall Recognition pairs might be better reealled and reeognized than the Cue AssociMean Mean ation High Low High Low Lo-Lo word pairs simply beeause the Hi-Hi target words are remembered better than the Lo-Lo targets. In order Low .39 .68 .54 .61 .65 .69 High .47 .65 .79 to test for this potential artifaet, a brief experiment .56 .82 .76 Mean .43 .67 .69 .76 was conducted in which the 48 target words were shown as isolated units, that is, not in the presence of the cue Note -Recall refers to the probability of eorreet recall; recogniwords, to an independent group of 10 subjects, who tion refers to the probability ofa hit. attempted to recognize and free recall as many of these Discussion words as possible. There were no differences in the The results of this experiment eorroborated the . hit rates among the four groups of words used in the main experiment [Hi-Hi = .69, Lo-Hi = .73, Hi-Lo = .68, expectation that the preexperimental association from Lo-Lo = .73; F(3,27) < 1], and there were no differ- cue to target would affect only eued reeall performance ences in the free recall scores [Hi-Hi = .33, Lo-Hi = .48, (Humphreys & Galbriath, 1975) and that the preexperimental association from target to cue would affect only Hi-Lo = .38, Lo-Lo = .42; F(3,27) = 2.7, P > .05] . A comparison between the recall (of the recall recognition. What appears most striking about these group) and recognition scores shows that recognition findings is not so much the differenee between recall [p(hit) = .72] was generally superior to recall [p(correct and recognition, but the similarity between them. In recall) .55] ; however, this was not necessarily the case both cases, high preexperimental association from the across all word pairs: For the Hi-Lo word pairs, recall retrieval eue to the eneoded trace led to better retention [pecorrect recall) = .68] was better than recognition than did low association. In the reeognition task, where [p(hit) = .61] . A finer grain analysis of recall vs, recogni- the retrieval cue was a nominal copy of the target, the tion differences comes from following the fate of an effeetive retrieval manipulation was the association individual item for a given subject across the two tests. from the target (to cue). In recall, where the retrieval The following analyses are based on the scores of those cue was a list cue, the effective retrieval manipulation 12 subjects in the recognition group who performed the was the association from the eue (to target). It appears recall test after having completed the recognition test. that both effects are alternate versions of the same For our purposes, the critical statistic is the prob ability general principle that memory for an event is related of recalling a target word when the cue word is supplied, to the ease of reconstructing the event from the retrieval given that the target word was not recognized in cue. Generally , this reconstruction operation has been the earlier test. For the four word-pair types, these interpreted as being a retrieval function. If we are to consistently apply this interpretation, we are led to the probabilities were .34 (15/44: Hi-Hi), .48 (29/60: Hi-Lo), .09 (3/32: Lo-Hi), and .05 (3/55: Lo-Lo). That eonclusion that retrieval is a component of both recall is, under appropria te recall conditions, when the and reeognition. The view that recall and recognition are similar in retrieval eue genera ted the target with high probability , nature and differ primarily on the basis of the informaa large proportion of words that were not recognized tion contained in their respective retrieval eues has been could be recalled. When the recall scores of these 12 subjects (recogni- endorsed recently by several investigators (e.g., Ftexser tion group) are conditionalized on correct earlier & Tulving, 1978; Lockhart et al., 1976; Rabinowitz, Mandler, & Patterson, 1977), all of whom maintain that recognition, the probabilities of recall for the four successful memory in both paradigms is related to the word-pair types are .80 (80/100: Hi-Hi), .86 (72/84: degree to whieh the retrieval eue accesses or overlaps the Hi-Lo), .57 (64/112: Lo-Hi), and .52 (46/89: Lo-Lo). Two sets of relations should be noted when comparing information in the ariginally eneoded trace. Two related these scores to the previously mentioned eonditional expectations emerge from this view. First, as previously scores. First, for eaeh word-pair type, the level of mentioned, if the information in the contextual and copy cues are independent, then the recallability reeall is higher when eonditionalized on recognition and recognizability of an event ean be manipulated success than when conditionalized on reeognition failure, indicating some general faetar that eontributes separably. Second, if the encoded trace being accessed to both recall and recognition. Second, reeall is relatively is identical for the two tasks, an experimental manipulagood for the Hi-Hi and Hi-Lo word pairs and poor for tion that alters the encoded traee should have eorrelated the Lo-Hi and Lo-Lo word pairs whether conditionalized effects on recall and recognition. Any learning manipuon recognition success or reeognition failure , indieating lation that makes the eneoded event gene rally more some factor that is speeifie to the reeall test. accessiblc , that is, from a variety of retrieval cues,
=
228
FISHER
should support better reeall and reeognition. This may be eonceptualized in any of several ways, depending upon one's theoretieal framework of the meehanism responsible for "informational overlap." Suppose, for example, that informational overlap between the eneoded traee and the retrieval eue is direetly related to the number of features shared by the two such that the greater the number of eomrnon features, the more likely is reeolleetion of the original event (Flexser & Tulving, 1978). A leaming manipulation that inereases the number of eneoded features of the original event should thereby make it more retrievable when any of several (e.g., reeall and reeognition) eues are used. In Experiment 1 no attempt was made to implement an experimental manipulation that would have similar effeets on reeall and reeognition, so this suggestion is not direetly testable from the data of the first experiment. However, an analysis of the eonditional reeall scores tends to eonfirm this idea: For all four word-pair types, eued reeall of the target word was higher when eonditionalized on sueeessful reeognition of the target than when eonditionalized on failure to reeognize the target. Now, one may wish to argue that this positive correlation between reeall and reeognition scores refleets in part the contribution of some "artifaet" like item seleetion (there is some inherent eharaeteristie of eertain word pairs that makes them relatively easy to learn) or some attentional faetor (e.g., the subjeet pays more attention to some word pairs than others and thus leams them better). However, this is in perfeet agreement with the suggestion presented here; namely, the eneoded traee that the learner attempts to aecess is identieal for reeall and reeognition tests. As such, whatever makes the trace generally more aeeessible, whether it is a property of the item to be learned or a eharaeteristie of the leamer, will benefit both reeall and reeognition performance (Flexser & Tulving, 1978; Rabinowitz et al., 1977). The results of Experiment 1 partially support the suggestion that two separate classes of experimental manipulations exist, one of whieh has separable effeets on reeall and reeognition, and the other eorrelated effeets. In order to more appropriately test this idea, in Experiment 2 both classes of experimental manipulations were direetly implemented. First, preexperimental target-to-eue and cue-to-target associations were again expeeted to have separable effeets on reeognition and reeall. Seeond, leaming instructions, whether to learn the pair of words by repetition or by forming a meaningful relation between the two words, were expeeted to yield eorrelated effeets on reeall and reeognition. Meaningful learning was expeeted to yield a rieher, more elaborate traee and thereby support better reeall and reeognition (see Fisher & Craik, Note 1). EXPERIMENT 2 Method
Subjeets and Design. The basic design of Experiment 2 was similar to that of Experiment I with the exeeption that
an additional variable, learning instruetions, was introdueed. Forty-eight male and female Erindale College students participated in this experiment. Twenty-four subjects were instrueted to learn the word pairs by repeating the pairs to themselves (repetition), and 24 subjeets were instrueted to leam the word pairs by ereating meaningful relations between the two words (meaningful). Before the word pairs were presented, the experimenter suggested various types of meaningful relations that the subjeets might use (e.g., making up phrases or sentences containing the two words, generating interaetive images of the two words, finding categorical relationships or words that are eommon associates of both words in the pair). Procedure, All subjeets were told that they would see aseries of word pairs and that they should study these pairs as units, since they would be later tested on them. The subjects were not told whieh member of the pair would be the test item. The same set of 54 stimulus eards that was used in Experiment I was presented at a 6-see rate. After the last word pair was presented, all the subjeets received a eued recall and a recognition test. Test order was eounterbalanced aeross subjeets. The target words, retrieval eues, and foils were identical to those used in Experiment 1. In the reeognition test, the subjects were given unlimited time and were perrnitted to indicate as many "old" items as they wished. In the eued reeall test, the subjeets were permitted 7.5 sec per retrieval eue to write the appropriate response.
Results and Discussion A preliminary analysis showed no main effeet of test order, whether reeall preeeded or followed the reeognition test (F< 1), and no reliable interaction between order and test (reeall vs. reeognition) [F(l ,44) = 3.00, P > .05]. Unless otherwise indicated, the following analyses are based upon the scores eollapsed aeross the two test orders. The effeets of preexperimental eue-to-target and target-to-cue associations on reeall and reeognition ean be seen in Table 3; they closely resemble the pattern of data in Experiment 1. Recall performance was directly related to cue-to-target association, whereas reeognition performance was directly related to target-to-cue association. The differenee between the probabilities of eorreet reeall for high and low cue-to-target associations (.70 vs. .44) was statistieally reliable [F(1 ,46) = 178.6, P < .01] , as was the differenee between the probabilities of a hit (reeognition) for high and low target-to-eue assoeiations (.76 vs. .59) [F(I,46) = 95.5, p