Context Effects and the use of Relational Information in Recognition Memory

Context Effects and the use of Relational Information in Recognition Memory

Memory and Control of Action Richard A.Magil1 (ed.) 0 North-HollandPublishingCompany,1983 323 CONTEXT EFFECTS AND THE USE OF RELATIONAL INFORMATION ...

1MB Sizes 1 Downloads 69 Views

Memory and Control of Action Richard A.Magil1 (ed.) 0 North-HollandPublishingCompany,1983

323

CONTEXT EFFECTS AND THE USE OF RELATIONAL INFORMATION IN RECOGNITION MEMORY Robert C. Mathews, Laurie Heffernan, & Randell Elkins Louisiana State University

Studies of context effects on recognition memory are examined in relation to retrieval differences in recall versus recognition memory. Support is provided for the position that some episodic on a information is automatically accessed recognition test, but that additional, more precise contextual information requires effortful retrieval processes. It is suggested that two types of context effects on recognition memory exist in the literature, and which type of effect is obtained depends on subjects' strategies for using relational information from memory.

An underlying theme behind much of the interest in context effects on

recognition memory concerns the relative speed and automaticity of retrieval of information from memory on a recognition as compared to a recall task. For example, try to recall what you had for dinner last Tuesday. You will probably find that this task requires at least some time and effort. Perhaps it was necessary to first reconstruct where you were for dinner on that day and with whom you ate, before you are finally able to recall the main course and, subsequently, the accompanying dishes.0n the other hand, consider a similar recognition type question: Did you eat spinach last Friday? Here, the answer probably becomes available to you quicker and with minimal effort-information particularly if you rarely eat spinach. It seems as if about your last encounter with spinach becomes accessible immediately upon seeing the word.

-

It is important to note that both of the above questions about last week's meals concern what Tulving (1972) calls episodic information: They are questions about particular temporally dated events (meals) in one's life. They are not questions about perceptual or semantic characteristics of the words. The interesting point is that some episodic information about events in which words or their referents occurred seems to become automatically accessible when the word is encountered again on a recognition test. Viewed in this light, studies of context effects on recognition memory clarify what or how much episodic information is immediately accessible through presentation of a word or "copy cue". The generally accepted

3 24

R.C. Mathews, L. Hejfernan and R. Elkins

view about retrieval of episodic information on a recognition test may be summarized as follows: (a) some (but not all) of the episodic information about a word is immediately and automatically available from memory upon presentation of a copy of the word. (b) Additional episodic information from memory may be obtained from the copy cue through effortful retrieval processes which are similar to retrieval processes used in recall (Atkinson & Juola, 1974; Mandler, Pearlstone, & Koopmans, 1969; Robinowitz, Mandler, & Borsalou, 1 9 7 7 ) . (c) Finally, not all of the episodic information about a word is accessible through the copy cue. Other cues, such as additional contextual items associated with the same event, may provide access to additional information about the word which was not accessible through the copy cue (Thomson, 1972; Tulving & Thomson, 1 9 7 1 ) . Thus, context effects or improvements in recognition memory when contextual items are provided in conjunction with a copy cue identify information in memory which is not accessible through a copy cue alone

.

Dual code theory of recognition memory, (Mandler, Pearlstone, & Koopmans, 1969; Robinowitz, Mandler, & Borsalou, 1977) extends this general framework into a theory of recognition memory. The theory proposes that there are two distinct sources of information from memory which are used in making recognition judgments about words. One of these, presentation codes, is immediately and automatically accessible upon presentation of the word or copy cue. The information contained in presentation codes includes phonological, perceptual, and general frequency information about a word. General frequency information gives one a general impression of how recently a word (or its referent) has occurred. This is episodic information of the crudest sort; it tells one whether or not a word occurred recently but tells nothing about the circumstances, context, or actual time of the events in which the word occurred. The second source of information from memory about a word is called the conceptual code. Conceptual codes are not automatically accessible upon presentation of a copy cue. They require the same effortful, t h e consuming, retrieval processes associated with a recall task. Once retrieved, however, the information contained in conceptual codes may include more precise contextual information about specific events in which the word occurred. Thus, conceptual codes may contain more precise episodic information than presentation codes. The exact nature of the episodic information contained in conceptual codes is thought to depend on the encoding associated with the study task or event in which the word occurred. Dual code theory also describes in detail how subjects use information from presentation and conceptual codes in making recognition judgments. For each to-be-recognized (tbr) item the subject first retrieves (automatically) the presentation code. From the information contained in the presentation code, the subject derives an impression of the recency of occurrence of the item. If this recency estimate exceeds the subject's general impression of the expected recency of list words (the subjects 'old' criterion) then he/she responds 'old' (identifies the tbr item as a list word). If the recency estimate for the tbr word is lower than the minimum expected recency for list words (the subjects 'new'

Use of Relational Information

325

criterion) then the subject responds 'new' (identifies the item as a lure). It is only when the recency estimate for a tbr item lies between the subject's new and old criteria that additional information will be sought from memory. When this happens, subjects attempt to retrieve the conceptual code(s) for the word. When the conceptual code is retrieved, any additional new episodic information contained in the code is used to decide whether to classify the item as old or new. The literature on context effects on recognition memory seems extremely complex and conflicting. However, when viewed in terms of dual code theory, a variety of different results can be integrated and explained. The next section of this paper attempts to review some of these studies and illustrate how some of the results can be explained in terms of dual code theory. Two Types of Context Effects on Recognition Memory Despite a rich literature on the subject, as recently as 1975 Slamecka questioned whether there are any true effects of context on recognition accuracy. Slamecka (1975) reviewed a number of studies claiming to have demonstrated effects of list context on recognition of words. When the results of these studies were transformed into measures of target-lure discriminability, the effects of context usually disappeared. Slamecka concluded that there is no convincing evidence that recognition accuracy is enhanced by intralist context. the Slamecka review suggests, part of the controversy in the literature reflects differences in what one is measuring. Many studies, including most of those reviewed by Slamecka (1975), find an increased hit rate for recognition of the target words when they are tested in the presence of list context (e.g., DaPolito, Barker, & Wiant, 1972; Thomson, 1972; Tulving & Thomson, 1971; Underwood, 1974). However, they also get a corresponding increase in false alarm rate when lures are tested in the same context. Therefore, when recognition accuracy is measured by combining data concerning hit and false alarm rate (e.g., corrected recognition scores or signal detection measures) there is no net effect of context. As

However, to say context does not affect recognition accuracy is not to say that context has no effect on recognition memory. Clearly the above studies do show effects of context: Both targets and lures when presented in list context were more likely to be rated as 'old' items from the list. Dismissing this type of effect as "merely" a shift in subject's criterion for rating items as old, is to overlook a possibly interesting phenomenon. In addition to the above type of context effect on recognition, the literature also contains several instances of "true" effects of context on recognition accuracy. Although these effects do not seem to be rare, they do seem to occur under a more limited set of conditions. Underwood and Humphreys (1979) divide studies of context effects into three general types: Studies of (a) Context addition involve adding

326

R.C. Mathews, L. Heffernan and R . Elkins

contextual cues on the recognition test which were not on the study list. (b) Context deletion studies involve deleting cues on the recognition test which were present on the study list; and (c) Context substitution studies involve substituting new cues on the recognition test in place of old list cues. Using this classification scheme, one usually finds the strongest effects of context on recognition accuracy in studies using context substitution (e.g., Davis, Lockhart, & Thomson, 1972, Exp. 1 & 2; DaPolito, Barker, & Wiant, 1972; Humphreys, 1976; Hunt, 1975; Underwood, 1974). Underwood and Humphreys (1979) conclude that there are no effects of context addition on recognition accuracy. Finally, there are several studies reporting generally small but reliable effects of context deletion on recognition accuracy. But this type of context effect seems to occur under the most limited and not completely understood sets of conditions. Context deletion effects on recognition accuracy have been found in some studies in which the instructions or encoding task seems to emphasize relational encoding (e.g., Davis, Lockhart, & Thomson, 1972, Exp. 3; Humphreys, 1976; Underwood, 1976; Winograd, Karchmer, & Russell, 1971), and in some studies in which the context was semantically meaningful (e.g., semantically related words in Marcel & Steel, 1973 and sentences in Baker & Santa, 1977). Thus, the literature seems to contain two very different types of context effect on recognition memory: Type-A effects in which both hit rate and false alarm rate are increased when an item is tested in the presence of 'old' list context, and type-B effects or true effects of context on recognition accuracy. In type-B effects the hit rate is increased in the presence of context and the false alarm is only slightly affected or in some cases it decreases in the presence of context. The dual code theory of recognition memory (Mandler, Pearlstone, & Koopmans, 1969; Rabinowitz, Mandler, & Barsalou, 1977) appears very useful in explaining these two types of context effect. Type-A effects of context (increased hit and false alarm rate) occur when subjects make most of their recognition decisions based solely on presentation codes. This is likely to happen when there is a large discrepancy in familiarity of targets and lures on the recognition test or when there is little contextual information in the conceptual codes (e.g., when the encoding task requires the encoding of little relational information). The only necessary additional assumption one needs to add to dual code theory in order to predict type-A context effects is that the familiarity of the context items affects the familiarity estimate of the tbr item. That is both targets and lures are subjectively estimated as more familiar in the presence of a familiar context item, increasing the probability that both (targets and lures) will be rated as 'old' on the basis of presentation codes alone. The type-B, context effect, in which recognition accuracy is enhanced by context, is likely to occur when: (a) Subjects retrieve the conceptual codes of a large number of tbr items during the recognition test, and (b) the conceptual codes contain information which is useful in making recognition decisions. The latter condition is most likely to occur when the encoding task during presentation of the list required the

Use of Relational Information

327

encoding of relational information (e.g., Humphreys, 1976) and the context is at least somewhat meaningful (e.g., as when sentences were used as context (e.g., Baker & Santa, 1977). Given the theoretical analysis of context effects presented so for, whether one obtains a type-A or B context effect should be greatly dependent on subject's retrieval strategy: Any variable which would increase the probability of subjects retrieving conceptual codes of the tbr words on the recognition test should increase the chance of getting a type-B context effect. Next we will report three context deletion experiments attempting to induce use of conceptual codes, and thereby produce a type-B context effect. These experiments were conducted by the first author and Laurie Olson. Three Context Deletion Experiments General Method The basic method common to all three experiments included a semantic encoding task in which subjects made semantic judgments about triplets of words followed by a recognition test for words from the triplets. Recognition judgments were made with the words tested by themselves and with one or both other members of a triplet present as context. Experiments 1 and 2 employ a novel sequential-testing procedure in which a word is tested first without context, then immediately afterwards with one context item present, and finally a third recognition judgment is made with both contextual items present. The experimental question being addressed is to what degree the addition of contextual information improves recognition accuracy above the level associated with the copy cue alone (a type-B context effect). Although this method confounds l e v e l s of context (none, one item or both) with judgment order (first, second, third judgment), it is probably a much more sensitive test for recognition improvement than conventional methods because: (a) the recognition judgments with and without context are made in close temporal contiguity. Thus the large variability n recognition judgments for a single item on successive trials (e.g., see Rabinowitz, Mandler, & Barsalou, 1977, Exp. 4 ) will not mask the context effects. And, (b) this procedure makes it obvious to the subject that he/she should to utilize contextual information to improve subsequent recognition decisions. Since the retrieval of conceptual codes is thought to be at least partly under subjects' control this procedure should help insure that subjects will use the conceptual codes. Experiment 3 uses the more traditional between-subject design in which subjects made only one recognition judgment about each tbr word in the presence of just one level of contextual cues. The Encoding Task. A triplet semantic judgment task was selected because it is known to foster relational encoding (cf. Mathews, 1977; Mathews, Maples, & Elkins, 1981), and because the use of triplets allows testing the effects of more than just one item as context. In this task subjects are given a list of triplets, each consisting of a phrase defining a semantic concept (e.g., ON A FARM) and a pair of nouns. In referring to members of a triplet, I will refer to the phrase as the Standard concept and the two nouns as the alternatives. Subjects are

328

R.C. Mathews, L. Heffernan and R. Elkins

instructed to place a number (0, 1, or 2) next to each triplet indicating how many of the alternatives are instances of the standard concept. The triplets are divided into four conditions, based on the relation of the alternatives to the standard concept. Triplets in which both alternatives are instances of the concept (e.g., ON A FARM, cow, plow) are classified as members of the similarity condition. The negative attribute condition consists of triplets in which neither alternative is an instance of the standard concept (e.g., PETROLEUM PRODUCT, circus, jungle). The contrast condition consists of all triplets in which one (but not both) alternatives are instances of the standard concept (e.g., FOR KILLING INSECTS, leukemia, fly swatter). These triplets are subdivided into two conditions depending on which alternative is the target on the recognition test. In the contrast-positive condition, the target word an instance of the concept (e.g., fly swatter), and in the contrast-negative condition, the target is the alternative which is not an instance of the concept (e.g., leukemia). The recognition test. A l l recognition judgments were made on a four point scale (1 = definitely in the list, 2 = think it was in the list, 3 = think it was not in the list, and 4 = sure it was not in the list). Half of the to-be-recognized items on the recognition test were targets (list words) and half were lures. Lures were judged in the same context (triplet members) as the targets they replaced and they were selected so that they had a similar semantic relation to the context items as that of the target (e.g., aspirin was the lure for coughdrops in the triplet antibiotics. PRESCRIPTION NEEDED, coughdrops, _____ Experiment 1 Method Procedure. Subjects were given a list containing the triplets typed in three columns across the page. The phrase defining the standard concept was typed in the first column and the two alternatives were randomly assigned to the remaining two columns. Across from each triplet was a space for the subject's response (0, 1, or 2) indicating the number of alternatives which are instances of the standard concept. Subjects were told to take their time in making their judgments, and they were allowed to complete the task at their own rate. They were not told that their memory for the words was going to be tested in the second part of the experiment. Three ring binders in which each full page was preceded by two partial cover pages were used for the recognition test. When both cover pages were down, only one word (the target or lure) was visible. A s each cover page was turned over, one additional context item was exposed. Subjects were allowed to proceed through the binders at their own speed, making three successive recognition judgments about each tbr item. The first rating was to be made with both cover pages down, and one cover page was to be turned over for each additional judgment. Subjects were not allowed to change prior ratings, but they were encouraged to revise successive ratings based on information gained from the context. It was

Use of Relational Information

3 29

made clear to the subjects that each recognition judgment was only about the tbr word, that all context items were from the list of triplets, and that the context items might help them decide whether or not the tbr word was also from the list. Particular target and lure pairs alternated on two forms of the recognition test given to different subjects so that across all subjects each triplet was tested with a target and with a lure equally often. Materials. Each list used in the semantic judgment task contained 64 triplets, 16 from each of the four triplet conditions. Three of the 64 sets of items used to construct the lists are shown in Table 1. Four different lists of triplets were formed by combining one of the four standard concepts (column 1 of Table 1) with the target and other alternative (columns 2 and 3). Then an additional four lists of triplets were formed by switching the targets with the lures (in parentheses in Table 1) in each triplet. Thus a total of eight different lists were constructed, and across all eight lists, each target occurred in each encoding condition equally often. Subjects. A total of 56 Louisiana State University undergraduates enrolled in introductory psychology participated in the experiment for extra credit. Subjects were tested in small groups of 1 to 6 in an experimental session lasting about 30 minutes. Results Since the encoding conditions are defined in terms of an assumed pattern of semantic relations between the alternatives and the standard concept, data for which a subject's response on the semantic judgment task indicated a different pattern ( e . g . , a response of "2'' for a triplet in the contrast condition) were excluded from the recognition analyses. The number of triplets for which data were excluded for a particular subject ranged from 0 to 15 with a mean of 5.8. All findings reported as being reliable in these experiments are significant at or beyond the .05 level. The mean recognition ratings on the four point scale, combined across encoding conditions (all of which showed the same pattern) were: 1.53 for targets without context, 1.35 with the standard as context, 1.47 with the alternative as context, and 1.38 with both context items exposed. The mean ratings for lures were: 3.32, 3.01, 3.04, and 3.05 in the same four conditions, respectively. Thus, both targets and lures were rated lower (more likely to have been in the list) in the presence of context, Each subject's recognition ratings were converted to the measure (the area measure of signal detection theory) of target-lure discriminability using the procedure outlined by Brown (1976). This measure varies from 0 to 1 with .5 representing chance or no discrimination between targets and lures. Table 2 presents mean & scores as a function of encoding conditions and context for all three experiments.

R.C. Mathews, L. Heffernan and R. Elkins

330

Table 1 Samples of Materials Used to Form the Triplets

Encoding Condition

Standard

Target

Alternative

Tractor (plow)

cow

Coughdrops (aspirin)

Antibiotics

Hat (cap)

Shirt

Set 1 Similarity Contrast-positive Contrast-negative Negative attribute

On a farm Farm implement Gives milk Illegal Set 2

Similarity Contrast-positive Contrast-negative Negative attribute

Medicine No prescription Prescription needed Temporary structure Set 3

Similarity Contrast-positive Contrast-negative Negative attribute

Clothing Worn on head Worn on chest Breathes

Note: Lures are shown in parentheses under target they replaced on the recognition test. It is apparent from the mean & scores shown in Table 2 that context actually impaired target-lure discrimination in Experiment 1. All of the means with context (columns 2, 3 , and 4) are lower than the corresponding means without context (column 1). An analysis of variance revealed a significant main effect of context, F(3, 825) = 7.47, MS = .019. There was also a main effect of encoding conditioz F ( 3 , 8 2 5 ) = 6.79, &M = .019. The differences among the encoding conditions reflect the same pattern obtained when these conditions were compared in recall (Mathews, 1977), namely, higher scores in the contrast and similarity conditions than in the negative-attribute condition. It is also worth noting that negative targets, those which are not instances of the standard concept (those in the negative-attribute and contrast-negative conditions), were recognized less well than were positive targets (those in the contrast-positive and similarity conditions). The difference between the two contrast conditions (contrast-positive versus contrast negative) was also significant,1(1,825) = 8.51, = .019.

Use of Relational Information

33 1

Table 2 Mean & Score for Target-Lure Discriminability as a Function of Encoding Condition and Context

Encoding Condition No Context

Context Alternative

Standard Both Items

Experiment 1 Negative-attribute Similarity Contrast-negative Contrast-positive

.831 .872 .849 .896

.775 .804 .831

.El1 .El8 .821 .879

.785 .El7 .806 .825

.697 .790 .7 20 .790

.740 .738 .803 .8 1 6

.755 .772 .776 .810

.7 30

.744 .785 .858 .895

.722 .831 .839 .792

.827

Experiment 2 Negative-attribute Similarity Contrast-negative Contrast-positive

.697 .729 .691 .747

Experiment 3 Negative-attribute Similarity Contrast-negative Contrast-positive

.732 ,791 .793 .798

.912

.781 .El1

Note: Each mean in Experiments 1 and 2 is based on 5 6 subject means and in Experiment 3 on 1 4 subject means. Standard error of the mean within cells ranged from .01 - .04 in the three experiments.

Discussion The main finding of this experiment is that context had the exact opposite effect on recognition accuracy from that we had anticipated: The presence of either one or both additional members of a triplet as context impaired target-lure discriminability relative to the no-context condition. This result appears to be an exaggerated example of a type-A context effect in which both hit rate and false alarm rate are increased when context is present (e.g., DaPolito, Barker, & Wiant, 1972; Tulving & Thomson, 1 9 7 1 ) . Whether the net effect of this type of context effect on recognition accuracy is positive, negative, or zero would depend on whether hit or false alarm rate increased the most. Theoretically,

332

R.C. Mathews, L. Heffernan and R. Elkins

whether the hit or false alarm rate is most affected by context would depend on the position of the signal (target) and noise (lure) distributions with respect to the 'old' criterion. In this experiment it was the false alarm rate which increased most resulting in a net impairment of recognition accuracy by context. In terms of dual code theory (e.g., Rabinowitz, Mandler, & Barsolou, this result suggests that subjects in this experiment, in spite of our efforts to encourage retrieval of conceptual codes, relied most on presentation codes in making their recognition judgments. Presumably, the familiarity of the context items increased the familiarity estimates of both targets and lures when presented in context. Thus, both targets and lures were more likely to exceed the 'old' criterion when presented in context. 1977),

If the above analysis of the results of Experiment 1 is correct, subjects may have had useful contextual information in their memories (conceptual codes) that they did not use on the recognition test. The question remains: How can we get subjects to retrieve and use the contextual information available in their contextual codes to enhance their recognition accuracy? In Experiment 2, we try another approach: In this experiment we attempted to make the use of presentation codes more difficult. In Experiment 2 both targets and lures occurred recently during the experimental session on the recognition test subjects had to discriminate triplet members from other words (lures) which occurred in another task prior to presentation of the triplet list. Experiment 2 If the proposed explanation for the impairment of recognition by context in Experiment 1 is correct, it should be possible to reverse the direction of context effects, (from negative to positive) by inducing subjects to rely more on conceptual rather than presentation codes in making their recognition decisions. Experiment 2 attempts this reversal by adding another task to the design of Experiment 1 in which both targets and lures occur. This new, word-classification task becomes the first task performed by subjects in Experiment 2 . It is followed by the same semantic judgment triplets and recognition test for words from the semantic judgment task used in Experiment 1 . The difference here is that, since both targets and lures occurred recently during the experimental session (in the word classification task), both sets of items will have recent general familiarity estimates associated with them on the recognition test. Since presentation codes should be of little use in discriminating targets from lures, subjects should be forced to rely more on conceptual codes in making their recognition judgments. Therefore, it was predicted that context would facilitate recognition accuracy in Experiment 2.

Use of Relational Information

333

Method Procedure. Experiment 2 was an exact replication of Experiment 1 with the additional word classification task added to the experimental procedure. Subjects in Experiment 2 performed the word classification task first, then the semantic judgment task, then their recognition memory for words from the semantic judgment task was tested. The sole purpose of the word classification task was to provide an opportunity to present the targets and lures from the recognition test before the triplets of the semantic judgment task were presented. For the word classification task subjects were given a list containing a random arrangement of the 6 4 targets plus the 64 lures from both forms of the recognition test ( s e e Materials Section of Experiment 1). They were asked to place an "A" or an "I" next to each word on the list indicating whether it referred to an animate or an inanimate object, respectively. As soon as subjects finished this task, the lists were collected and the semantic judgment task was begun. From this point on the procedure was identical to that of Experiment 1. Results As in Experiment 1 data from triplets which were misclassified on the semantic judgment task were excluded from the recognition analyses. The mean recognition ratings for targets, combined across encoding conditions, were: 1.61 with no context, 1 . 4 2 with the standard as context, 1.57 with the alternative as context, and 1 . 3 9 with both context items present. For lures the corresponding means were 2 . 4 7 , Thus, unlike the results of 2 . 6 7 , 2 . 6 2 , and 2 . 7 9 , respectively. Experiment 1, ratings for targets decreased in the presence of context and ratings for lures increased; suggesting that recognition accuracy was facilitated by context in this experiment. The pattern of 5 scores for target-lure discrimination in Experiment 2 (see Table 2 ) confirms that recognition accuracy was facilitated by context. In an analysis of variance performed on the & scores, the main effect of context was significant, F ( 3 , 8 2 5 ) = 9 . 3 4 , M & = .020. Except with the alternative as context in the for one tie (.697 negative-attribute condition), all of the mean & scores are greater with context (columns 2 , 3 , & 4 of Table 1) than without (column 1). Another interesting observation concerning the pattern of & scores in Experiment 2 is that the effects of the two context items (standard and alternative) do not appear to be additive. That is, while in most cases each item facilitates recognition compared to the no-context condition, the presence of both items as context does not generally result in better recognition than both single-item context conditions. In fact, within encoding conditions (rows of Table l), the highest 5 score falls in the both context column only in the negative attribute condition. Discussion As predicted, the effect of context on recognition accuracy was reversed (from impairment in Experiment 1 to facilitation in Experiment 2) without changing the materials, encoding task, or the nature of the

334

R.C. Mathews, L. Heffernan and R. Elkins

recognition test used in Experiment 1. Thus, it is likely that subjects in Experiment 1 did have contextual information about the triplets in their memories which could have been used to improve their recognition accuracy. However, we could not get subjects to use this information (in Experiment 1) until we decreased the usefulness of presentation codes by insuring that both targets and lures were highly familiar on the recognition test (Experiment 2). The sequential testing procedure used in these two experiments has the advantages of within-subject, within-item comparisons made in close temporal contiguity. However, its major disadvantage is that addition of context is confounded with sequence of judgment. Therefore, we will report a third experiment using a between subject design. In Experiment 3 , subjects made only one recognition decision about each tbr item under just one level of context, Experiment 3 Method The method of Experiment 3 was identical to that of Experiment 2 except that each subject made just one recognition judgment about each tbr item. Subjects in this experiment were divided into four groups, each group making all their recognition judgments under one context condition: One group made all their judgments with no context, another with the standard concepts as context, a third with the alternatives as context, and the fourth group with both context items present. A s in the previous experiments 56 Louisiana State University undergraduates (14 in each group) served as subjects. Results The mean & scores from Experiment 3 are shown in Table 2. While the main effect of context in this experiment was not significant, there were clear instances of facilitory effects of context as reflected in the significant context by encoding condition interaction, >(9,156) = 2.04, MSe = .013. Furthermore, the pattern of & scores in Experiment 3 was very similar to that of Experiment 2. Taking the differences between respective context and no-context conditions as a measure of the size of the context effects, relatively large effects were obtained in both experiments with the alternative as context in the similarity condition (differences in mean score of .061 and .121 in the two experiments, respectively) and with the standard as context in both contrast conditions (differences ranging from .065 - .112). Little or no facilitory effects were found with the standard as context in the similarity condition (differences of .009 and .006, respectively). Conclusions from Experiments 1-3 The results of these experiments are consistent with the notion that two types of context effects on recognition memory occur: Experiment 1 produced a type-A context effect in which both targets and lures are rated as more likely to have been in the list when presented in context. Experiments 2 and 3 produced a type-B effect of context on recognition accuracy. The combined strategy of using both the pre-familiarization

Use of Relational Information

335

task in which both targets and lures occurred prior to presentation of the to-be-remembered list and the use of the sequential testing procedure (Experiment 2 ) produced the strongest facilitory effects of context on recognition. Experiment 3 which used only the prefamiliarization task produced a similar but weaker pattern of context effects to those found in Experiment 2. The results of the three experiments are also consistent with predictions derived from dual code theory (Mandler, Pearlstone, & Koopmans, 1969; Rabinowitz, Mandler, & Barsalou, 1977). Forcing subjects to rely more on conceptual codes by decreasing familiarity differences between targets and lures in Experiments 2 and 3 successfully reversed the directions of the effect of context on recognition accuracy from that obtained in Experiment 1 (impairment). The negative effect of context on recognition accuracy in Experiment 1 is of interest in its own right. We view this results as an exaggerated example of a type-A context effect in which both targets and lures are rated as more likely to be list members when presented in list context. The occurrence of this type of context effect suggests that the familiarity estimate for an item (used to decide if the item can be classified as old or new based on its presentation code alone) is influenced by context. That is, the familiarity of the context item itself seems to spill over, increasing the familiarity estimates for both targets and lures presented in that context. One additional curious, and new finding from these experiments concerns the lack of additivity of facilitory effects of the two triplet members as context. It was often the case that both single items from a triplet (columns 2 and 3 of Table 2) facilitated recognition of the target item as compared to the no context condition (column 1). However recognition accuracy was best with both items present as context (column 4 ) only in 1 of 8 relevant comparisons in Experiments 2 and 3 . Since virtually all current models of recognition memory suggest that recognition accuracy depends on the degree of match between the encoding of a test item and the item's original encoding in the list (e.g., Anderson & Bower, 1974; Glass, Holyoak, & Santa, 1979; Kintsch, 1974) this finding is very odd. It would seem logical that the match between original and test encoding would be maximized by presenting the complete triplet context on the recognition test. Further studies are needed in which the original items are presented in units of 3 or more (e.g., triplets) so that this result may be replicated and explored. Finally, it is clear from these results that different members of a triplet contribute differently to the obtained context effect depending on triplet type (similarity, contrast, or negative attribute). In Experiments 2 and 3 the alternative item contributed most to recognition improvement in the similarity condition; whereas the standard contributed most in the contrast condition. This pattern of results appears to defy simple explanations in terms of encoding task, semantic congruity (e.g., Baker & Santa, 1977; Schulman, 1974), or in terms of the tagging of particular associations during study (e.g., Anderson & Bower, 1973). Instead, the potency of particular items as context seems to be dependent on some type of interaction between the encoding task

336

R.C. Mathews, L. Heffernan and R . Elkins

and the semantic-associative structure. Perhaps this interaction is related to what Begg (1978) refers to as "distinctive similarity of the cue-trace relation relative to other traces and cues." Relational Information and Its Use in Recognition Memory far this paper has taken the position that there is often contextual information within a subject's memory (conceptual codes) which--if retrieved--can facilitate recognition accuracy. The remainder of this paper concerns the nature of this information and further findings concerning strategies subjects have for using this type of information in certain memory tasks. Presumably not all information within conceptual codes is useful in distinguishing targets from lures. Facilitory effects of context on recognition accuracy require situations in which information in memory about one item (the context) facilitates recognition of another item as a list member. Thus, type-B context effects would seem to require information within conceptual codes which links items as having occurred together on the presentation list. Such information about the co-occurence of list items seems equivalent to what some authors speak of as associative (e.g., underwood, 1974) or relational information (e.g., Einstein & Hunt, 1980; Humphreys, 1976). So

Assuming that it is relational information within the conceptual codes which can facilitate recognition accuracy, the chain of events necessary to produce type-B content effects grows longer: (a) Subjects must be induced to retrieve conceptual codes of the tbr words rather than relying on presentation codes. (b) The conceptual codes must contain relational information concerning the tbr item and its context. This is most likely to occur when the encoding task required relational processing (see Einstein & Hunt, 1980). Finally (c) subjects must make appropriate use of the relational information which was retrieved in making their recognition decisions. If relational information is responsible for type-B context effects, it should be possible to demonstrate the existence of relational information about an item and its context on a recognition task. This has been accomplished by Underwood (1974) and Humphreys (1976) in experiments in which subjects made recognition judgments about pairs of tbr items. A few important findings from each of these studies concerning the storage and use of relational information on a recognition task will be discussed below. The Underwood (1974) series consists of 9 experiments. The general method used in these experiments involved study of list pairs followed by recognition of either single items (item recognition) OK item pairs (pair recognition). On the pair recognition tests subjects were asked to circle only test pairs in which both items were from the list. Test pairs included intact list pairs from the study list, rearranged pairs from the list (items from two different study pairs), and pairs containing one or more new (non-list) items. In tests of pair recognition (e.g., Exp. 31, subjects were more likely to circle intact than rearranged pairs. Recognition responses to test pairs were clearly sensitive to which items occurred together as study pairs, suggesting that relational information was stored and used on the recognition test.

Use o f Relational Information

331

Experiment 6 of the Underwood (1974) series examines pair recognition for a variety of types of prexperimentally related pairs (e.g., antonyms, conceptual associates, parallel associates, synonyms, and homonyms) and for unrelated pairs. This experiment produced a number of interesting results: overall (collapsed across types of related pairs) related pairs were not recognized more accurately than were the unrelated pairs. However, there were generally more hits false alarms on related pairs. This effect was especially pronounced with homonym pairs. Thus semantic relatedness of study pairs appears to result in a Type-A (increased hit and false alarm rate) effect on pair recognition rather than facilitating recognition accuracy. The series of experiments by Humphreys (1976) provides an even more direct test for the use of relational information in recognition memory. Humphreys used a paradigm similar to Underwood's (1974): subjects studied list pairs and were later given recognition test pairs consisting of intact, rearranged, and old-new pairs. In addition to recognition like Underwood (1974), Humphreys testing for item and also tested for relational recognition. On a test of relational recognition subjects were asked to identify (recognize) only intact list pairs. Subjects ability to discriminate intact from rearranged pairs on the relational recognition test is a direct test for the existence of relational information in memory; without such relational information, subjects discrimination of these pairs should be at a chance level. Humphrey's subjects were able to discriminate intact from rearranged pairs, demonstrating the existence of relational information in memory. Another finding from the Humphreys (1976) study which fits nicely with the dual code theory explanation of context effects concerns the criterion adopted by different subjects on the recognition text. According to dual code theory the higher the criterion one sets for responding 'old', the more likely one is to retrieve and use conceptual codes. This occurs because a higher 'old' criterion widens the gap between the 'old' and 'new' criteria, and it is only when the familiarity estimate for tbr words falls in this gap that subjects retrieve conceptual codes. In Experiment 2 of this series, Humphreys calculated the 'old' criterion used by each subject on the recognition test and used the criterion to divide his subjects into two groups, those having high or low 'old' criteria. As predicted by dual code theory the size of the type-B context effect increased for high-criterion subjects and decreased for low-criterion subjects

.

Based on these experiments comparing item, pair, and relational recognition, Humphreys (1976) drew two main conclusions about the use of relational information in recognition: (a) the use of relational information in recognition is at least partially under subjects control, and (b) subjects may not always use relational information to their best advantage on a recognition test. One final experiment will be reported which further examines the use of relational and item information in recognition memory. This experiment was conducted at LSU by the first author and Randel Elkins.

338

R.C. Mathews, L. Heffernan and R. Elkins Experiment 4

Method This experiment consisted of two parts: (a) First subjects were asked to classify pairs of words on a list as semantically related or not. (b) In part 2 of the experiment subjects were tested for relational recognition of list pairs and for item recognition of list words presented in test pairs. Materials. The study lists for part 1 of the experiment consisted of 112 pairs of words selected from the Battig and Montague (1969) categorization norms. Two lists were constructed, each containing 56 related pairs (exemplars of the same category) and 56 unrelated pairs (exemplars of different categories). The same words were used in both lists. Four exemplars of each category occurred in each list. These four exemplars occurred as two related pairs in one list and as members of four different unrelated pairs in the other list. A total of 16 different recognition lists were prepared, eight to be used with each of the two study lists. Each recognition list contained 28 intact pairs from the study lists, 28 rearranged pairs and 56 old-new pairs. New words used in old-new pairs were exemplars of the same category as the old word they replaced. Similarly, rearranged pairs maintained category membership of the words being exchanged. Thus, all test pairs (intact, rearranged, or old-new) for related study pairs were related through membership in the same category; and all test pairs for unrelated study pairs were unrelated. Across the 16 different recognition lists, each study pair was tested by an intact pair once, a rearranged pair once, and by an old-new pair twice. The same set of recognition lists were used for the relational and for the item recognition tests. On the item recognition tests, one word in each pair was underlined as the to-be-recognized item. Half of the old-new pairs had the old word underlined and half had the new word underlined. Every word on the recognition test served as the to-be-recognized word in one of the 16 versions of the recognition test. Procedure. Subjects were tested in small groups. In the first part of the experiment each subject was given a list of word pairs and was instructed to place a plus sign next to semantically related pairs and a minus sign next to unrelated pairs. When this task was completed, the lists were collected and the subjects were informed that their recognition memory for the words was going to be tested. Each subject took both the item- and relational-recognition tests, one immediately following the other. The order of the two tests was counterbalanced across subjects. Each subject received a different version of the recognition test for the relational- and for the item-recognition test. Subjects responded to each test pair on both the item and pair-recognition tests by placing a number ( 1 - 4 ) next to each pair. The numerical code for responding, which was printed on the top of every recognition test, was: 1 = sure in list, 2 = think in list, 3 = think not in list, 4 = sure not in list. On the relational recognition test subjects were told to respond "in list" (1 or 2) only when the pair was

Use of Relational Information

339

an intact list pair from part 1 of the experiment. On the item recognition test, the instructions told subjects that their recognition decision was only about the underlined word, but that the other word in a test pair might help them decide whether or not the underlined word was in the list. They were to respond "in list" (1 or 2) when they recognized the underlined word as a list item. Subjects. Subjects were 64 Louisiana State University undergraduates who participated in the experiment for credit in a psychology course. Results and Discussion Relational recognition. Two sets of R-scores were obtained for relational recognition. One is a measure of discrimination of intact from rearranged pairs and the other is a measure of discrimination of intact from old-new pairs. The first measure is regarded as a purer measure of relational recognition because, if the subject recognizes the new item in old-new pairs. this item information could be used to discriminate intact from old-new pairs. Mean R-scores for relational recognition are shown in Table 3 as a function of test order and semantic relatedness of the pair. Table 3 Mean R-Score as a Function of Test Order and Pair Relatedness

Measure

First Test Related Unrelated

Second Test Related Unrelated

Relational Recognition: Rearranged vs Intact Old-new vs Intact

.69 .82

.67 .78

.63 .77

.63 .71

.79 .80 .81

.73 .75 .69

.88 .86 .85

.67 .64 .63

Item Recognition: Intact Pair Rearranged Pair Old-new Pair

A 2 (pair relatedness) by 2 (lure type, rearranged or old-new) by 2 (test order) analysis of variance was conducted on the pair recognition data. This analysis revealed a significant effect of test order, F(1,248) = 17.77, M& = .011; pair relatedness, F(1,248 =.6.62, MS = .011; and lure type, F(1,248 = 82.96, = .Oli. The lure type bf pair relatedness interaction only approached significance, F(1,248) = 2.67, M& = .011, E = .lo. A post-ANOVA test of simple effects showed that the effect of semantic relatedness was restricted to discrimination

340

R.C. Mathews, L. Heffernan and R. Elkins

of intact from old-new pairs, F(1,248) = 8.31, MS = .011. As is apparent from the means in Table-4, semantic relatzess had no effect on subjects' ability to discriminate intact from rearranged = .337, &M = .011. pairs, J(1.248) Item recognition. The data from the item recognition test were transformed into R-scores for item recognition using the old-new pairs with the new word underlined as the lures. The mean R-scores for item recognition are shown in Table 3 as a function of test pair relatedness (related or unrelated), test pair type (intact pair, rearranged pair, or old-new pair with old word underlined), and test order (pair or item recognition first). An analysis of variance of the item recognition data showed a significant effect of pair relatedness F(1,372) = 119.47, MS = .017. Items from semantically related study pairs were recognfzed more accurately than were items from unrelated study pairs. The only other significant effect was a test order by relatedness interaction, F(1.372) = 31.09, = .017. This interaction reflects the fact that items from related pairs were recognized best when the item recognition test was first, while items from unrelated study pairs were recognized best when the relational recognition test was first ( s e e Table 3). Conclusions About the Use of Relation Information in Recognition Memory First, we will briefly summarize the conclusions from Experiment 4 concerning relational and item recognition. Then, we will consider conclusions about the use of relational information from a broader perspective. The conclusions from Experiment 4 concerning relational recognition are as follows: (a) Subjects did have sufficient relational information about the study pairs in memory to discriminate intact from rearranged pairs above the chance level. This was true despite the fact that the orienting task was very simple (indicate whether each study pair was semantically related or not), subjects did not know their memory--let alone relational memory--was going to be tested, and the recognition task was very difficult (discriminate intact from rearranged study pairs in which category membership was preserved). Thus, we conclude that subjects readily and easily store relational information about list pairs in memory, and they can effectively use relational information when the task demands it. (b) Subjects were better at discriminating intact from old-new than intact from rearranged pairs. Thus, it appears the information about the new item in old-new pairs was used to advantage on the relational recognition test. (c) Pair relatedness had no effect on discrimination of intact from rearranged pairs. Thus, relational information about which members of conceptually related pairs occurred together as study pairs in the list was preserved in memory no better than similar relational information about conceptually unrelated pairs. This finding is consistent with other results concerning recognition memory (e.g., Hermann & McLaughlin, 1973; Underwood, 1976),

Use of Relational Information

341

which suggest that storage of episodic relations among list items is relatively independent of semantic relations among the words. The main findings of Experiment 4 concerning item recognition were: (a) There was no effect of study-pair context on item recognition accuracy; that is, there was no type-B context effect on item recognition. Here is a situation where we know subjects had relational information about the study pairs in memory from their performance on the relational recognition test; and yet, apparently, this information was not used on the item recognition test. The dual-code theory explanation for this effect is that subjects relied primarily on presentation codes in making their item recognition decisions. Given this explanation is correct this finding is remarkable, particularly when the item recognition test was taken after the pair recognition test. In this case, subjects must have retrieved the conceptual codes of the words to perform the relational-recognition test just minutes before the item-recognition test. Yet, on the item recognition test they still responded to most the items on the basis of presentation codes alone. (b) There was a significant effect of semantic relatedness on item recognition. Items from semantically related pairs were better recognized than were items from unrelated pairs. This effect probably arises from the encoding task (i.e., it requires less processing to make decisions about unrelated pairs, e.g., Craik & Tulving, 1975; Mathews, Maples, & Elkins, 1981). The importance of this effect here is that it did not extend to relational recognition (of intact versus rearranged pairs). Thus, even though items in unrelated pairs were encoded at a lower level than were items in related pairs (as indicated by lower item recognition), unrelated and related list pairs were discriminated from rearranged list pairs equally well. This finding supports the notion that item and relational information are distinct sources of information in memory (e.g. Einstein & Hunt, 1980; Humphreys, 1976). On the whole, studies of relational information in recognition (Experiment 4 of this series, Humphreys, 1976; and Underwood, 1974) have shown that subjects readily encode relational information about co-occurring list items (e.g., study pairs). However, the use of relational information on a recognition test appears to be under subjects control (Humphreys, 1976); and subjects do not always seem to use available relational information to their best advantage (e.g., on an item recognition test). Dual code theory (Rabinowitz, Mandler, & Barsalou, 1977) offers at least a working hypothesis to explain why subjects so often fail to use available relational information on an item recognition test: Whenever the general familiarity of a tbr item seems sufficient to declare the item 'old', further information from memory (conceptual codes) is not sought. Furthermore, the results of Experiments 1-3 of the series suggest that the general familiarity of a tbr item is strongly influenced by the familiarity of the test context it occurs in.

R.C. Mathews, L. Heffernan and R. Elkins

342

References Anderson, J.R., & Bower, G. H. Human associative memory. New York: Wiley, 1973 Anderson, J. R., & Bower, G. H. A propositional theory of recognition memory. Memory & Cognition, 1974, 2, 406-412. Atkinson, R. C., & Juola, J. F. Search and decision processes in recognition memory. In D. H. Krantz, R. C. Atkinson, & P. Suppes (Eds.). Contemporary development in mathematical psychology. San Francisco: W. H. Freeman, 1974. Battig, W. F., & Montague, W. E. Category norms for verbal items in 56 categories: A replication and extension of the Connecticut category norms. Journal of Experimental Psychology Monograph, 1969, (3, pt 2). Baker, L., & Santa, J. L. Semantic integration and context. Memory & Cognition, 1977, 5, 151-154. Begg, I. Similarity and contrast in memory for relations. Memory & Cognition, 1978, 6, 509-517. Brown, J. An analysis of recognition and recall and of problems in their comparison. In J . Brown (Ed.), Recall and recognition. New York: Wiley, 1976. Craik, F. I. M., & Tulving, E. Depth of Processing and retention of words in episodic memory. Journal of Experimental Psychology: General, 1975, I, 268-294. DaPolito, F., Barker, D., & Wiant, J . The effects of contextual changes on component recognition. American Journal of Psychology, 1972, 85, 431-440. DavisFJ. C., Lockhart, R. S., & Thomson, D. M. Repetition and context effects in recognition memory. Journal of Experimental Psychology, 1972. 92. 96-102. ~Einstein, G. O., & Hunt, R. R. Levels of Processing and organization: Additive effects of individual-item and relational processing. Journal of Experimental Psychology: Human Learning and Memory, 1980, 5, 588-598. Glass, A. L., Holyoak, K. J., & Santa, 3 . L . Cognition. Reading, Mass: Addison Wesley, 1979. Herrmann, D.J., & McLaughlin, J. P. Effects of experimental and Dreexuerimental ornanization on recognition: Evidence for two storage systems in long-term memory. Journal of Experimental Psychology, 1973, 99, 174-179. Hunt. R. R. How similar are context effects in recognition and recall? Journal of Experimental Psychology: Human Learning & Memory, 1975, 1. 530-537. Humphfeys, M. S. Relational information and the context effect in recognition memory. Memory & Cognition, 1976, 4, 221-232. Kintsch, W. The representation of meaning in memory. New York: Wiley, 1974. Mandler, G., Pearlstone, Z., & Koopmans, H. S. Effects of organization and semantic similarity on recall and recognition. Journal of Verbal Learning and Verbal Behavior, 1969, 8, 410-423. Marcel, A., & Steel, R. G. Semantic cueing in recognition and recall. Quarterly Journal of Experimental Psychology, 1973, 25, 368-377. Mathews. R. C. Semantic -iudgments as encoding The effects - operations: of attention to particular semantic categories on the usefulness of

so,

I

.

.

-

Use of Relational Information

343

interitem relations in recall. Journal of Experimental Psychology: Human Learning and Memory, 1977, 1, 160-173. Mathews, R. C., Maples, R.C., & Elkins, R. Semantic judgments as encoding operations in recall: The encoding of task-relevant and task-irrelevant semantic attributes of words. Journal of General Psychology, 1981. 105,311-320. Rabinowitz, J. C., Mandler, G . , & Barsalou, L. W. Recognition failure: Another case of retrieval failure. Journal of Verbal Learning and Verbal Behavior, 1977, 16,639-663. Schulman, A.I. Memory for words recently classified. Memory & Cognition, 1974, 2,47-52. Slamecka, N. J. Intralist cueing of recognition. Journal of Verbal Learning and Verbal Behavior, 1975, 14,630-637. Thomson, D. M. Context effects in recognition memory. Journal of Verbal Learning and Verbal Behavior, 1972, 11,497-511. Tulving, E. Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press, 1972. Tulving, E., & Thomson, D. M. Retrieval processes in recognition memory: Effects of associate context. Journal of Experimental Psychology, 1971, 3, 116-124. Underwood, B. J. The role of association in recognition memory. Journal of Experimental Psychology, 1974, 102, 917-939. Underwood, B. J. Recognition memory for pairs of words as a function of associative context. Journal of Experimental Psychology: Human Learning and Memory, 1976, 2,404-412. Underwood, B. J., & Humphreys, M. Context change and the role of meaning in word recognition. American Journal of Psychology, 1979, 92, 577-609. Winoarad. E.. Karchmer. M.A., & Russell, I. S. Role of encoding unitization in cued recognition memory. Journal of Verbal Learning and Verbal Behavior, 1971, lo, 199-206. I

.