Transfer appropriate forgetting: The cue-dependent nature of retrieval-induced forgetting

Abstract

Retrieval-induced forgetting is the failure to recall a previously studied word following repeated retrieval of a related item. It has been argued that this is due to retrieval competition between practiced and unpracticed items, which results in inhibition of the non-recalled item, detectable with an independent cue at final test. Three experiments were conducted in which two cues were associated with a target item at encoding. All three studies demonstrated retrieval-induced forgetting when the same retrieval cue was present at practice and test, but not when the second encoding cue was used as an independent probe at final test. These data are not compatible with a general inhibitory account of retrieval-induced forgetting, but support a context-specific account of the phenomenon.

Introduction

In order to have a retrieval system that functions efficiently it is critical not only to recall what is desired, but to not recall what is unwanted. Given the content-addressable nature of human memory, this feat is quite an achievement. Consider, for example, answering the following question: “Where did you go on your last holiday?” Presumably, each of us has experienced a number of holidays that differ in their availability to recall. For example a trip abroad, lasting several weeks may be more memorable than a short break in our own country. Thus when we are asked this question, many potential memories may be activated, and we have to select between them in order to provide an answer to the question.

How does the memory system quickly and efficiently achieve this selection process, and what is the consequence of this? The present work focuses on one putative aspect of the selection process, namely retrieval inhibition. In answer to the question about our most recent holiday, numerous holiday memories might be potentially available to recall, but only one is the most recent. Thus, it might aid selection of the desired item if the activated memories that do not fit the requirements of the memory search are deactivated, whilst at the same time allowing the activation of the desired memory to accrue. Such mechanisms—of activation and competition between competing representations—are at the heart of many distributed models of human memory and language (e.g., McClelland & Rumelhart, 1986).

The particular focus of the present research is on the long-term consequences of this retrieval competition. It has been repeatedly shown that following retrieval competition, the non-retrieved item becomes less available to recall from long-term memory (Anderson & Bell, 2001; Anderson, Bjork, & Bjork, 1994; Anderson, Green, & McCulloch, 2000; Anderson & McCulloch, 1999; Ciranni & Shimamura, 1999; Koutstaal, Schacter, Johnson, & Galluccio, 1999; MacLeod, 2001; MacLeod & Macae, 2001; Macrae & MacLeod, 1999; Shaw III, Bjork, & Handal, 1995; Smith & Hunt, 2000). Further, it has been proposed that any explanation of this poorer recall performance requires an inhibitory component that acts upon the memory representation itself, rather than through any associative process such as unlearning (Anderson, 2003; Anderson et al., 2000; Anderson & Spellman, 1995; Levy & Anderson, 2002). Evidence for this assertion has come from research using the retrieval-induced forgetting paradigm.

The retrieval-induced forgetting paradigm was first introduced by Anderson et al. (1994). The paradigm has three critical phases. In the study phase, participants are asked to learn category cue—exemplar pairs for a later memory test. Critically, the list they see contains several members of each of a number of categories. So, they might see a list that includes the items FRUIT—orange, FRUIT—apple, TOOL—hammer, and TOOL—spanner amongst others.

The second phase is the retrieval practice phase, and is presented to the participants as a memory test for the material they have just studied. They are cued to recall some of the items from some of the categories, by means of a category plus word stem cue, such as “FRUIT—or____?” for orange. This stage is designed to elicit retrieval for a subset of items, which become designated the practiced (RP+) items. This creates a subset of items that come from the same categories that were not practiced (e.g., apple). These items are conventionally designated RP− items. Finally there are yet further items which are not practiced, but which are members of categories which were not practiced (e.g., hammer). These items are called either NRp (no retrieval practice) or U (unpracticed) items. We will use the latter term here.

After a 20 min filled delay the final stage of the paradigm is conducted, which is a category cued recall test for all the items that were studied. Anderson et al. (1994) demonstrated that the effect of repeatedly retrieving some members of a category from the category cue was to reduce the likelihood of recalling the unpracticed members of that category. That is, compared to the baseline of recall of U items, RP− items were less likely to be recalled. Thus, retrieving the item orange has the counterintuitive effect of making the related item apple less likely to be remembered 20 min later.

Anderson et al. (1994) argued that their data were supportive of an inhibitory account of the retrieval-induced forgetting paradigm. Of particular relevance to this argument were the effects on strong (highly available) and weak (less available) members of practiced categories. Across the three experiments reported, strong members of a category showed a 15% recall benefit when they were RP+ items and a 10% impairment when they acted as RP− items, relative to recall of baseline, U items. For weak members of a category, there was a 21% benefit for RP+ items, but no impairment for those items when they were RP− items (in fact there was 3% recall benefit). The presence of retrieval-induced forgetting for strong items occurred whether strong or weak items were practiced. Similarly, the lack of retrieval-induced forgetting for weak items was also independent of the strength of the practiced items. Anderson et al. argued that this pattern is inconsistent with strength dependent competition models, but instead favours an account based upon suppression of the non-practiced item. In this account, stronger category members get inhibited because they are more likely to compete for retrieval, whatever the strength of the practiced item. Conversely, weaker items are unlikely to compete for retrieval during practice, and so there is no need to inhibit them. Anderson et al. (1994) concluded that their results “..are consistent… with a process of active suppression, applied directly to competing items to the extent that those items interfere with task demands.” (p. 1080).

Whilst Anderson et al. (1994) argued for an inhibitory account of their findings, Anderson and Spellman (1995) acknowledged that there are several non-inhibitory accounts that could potentially explain those data. (For reviews of these non-inhibitory accounts see Anderson & Bjork, 1994; Anderson & Neely, 1996). For example, it is possible that the repeated retrieval practice of FRUIT—orange is likely to strengthen the association between the cue FRUIT and the exemplar orange, such that when FRUIT reappears as a cue, then orange is the most likely item to be recalled. Indeed, it is possible that orange is so strongly available that other items can no longer be recalled; in this model access to apple may be occluded by the repeated access of the item orange. Notice that the failure to recall apple in this model says nothing about the strength of the memory representation of apple, nor about the strength of the association between FRUIT and apple. The observed forgetting is caused merely by the strengthening of the FRUIT—orange link, together with a retrieval process that is resource limited. That is, only one item can be recalled at a time, and retrieval attempts will cease if the desired answer is not found quickly enough.

A different form of associative model based upon the principle of unlearning (Melton & Irwin, 1940) might also suffice. In such a model, the repeated retrieval of orange leads not only to the strengthening of the FRUIT—orange association, but also to the weakening of the association between FRUIT and apple. Thus when the cue FRUIT appears, apple is less likely to be cued, and so is forgotten. Note here, as before, the representation apple is not inhibited, since it is the FRUIT—apple link that is weakened, not the representation of apple itself.

Anderson and Spellman (1995) argued that the only way to distinguish between inhibitory and associative models of retrieval-induced forgetting is to use an independent retrieval cue at the final test stage (see also Levy & Anderson, 2002). They argued that all the non-inhibitory accounts share an assumption that the forgetting is due to a failure of the cue used at practice to elicit the target at test. Thus, the forgetting effect should be seen only if the same cue is used at the final test phase as was used in the practice phase. However, if retrieval-induced forgetting occurs because the non-retrieved item is inhibited at practice, then it follows that the forgetting should be seen with any cue, since it is the memory itself that is inhibited, rather than the link with a specific practice cue. This is the basis of the independent probe version of the retrieval-induced forgetting paradigm introduced by Anderson and Spellman (1995), and extended by Anderson et al. (2000).

In the independent probe technique, the essential elements of the retrieval-induced forgetting paradigm are retained, with the only modification used being the final test for the putatively inhibited item, which is cued by a different word to that used to elicit the practiced item. Two different forms of independent probe techniques have been used. In Anderson and Spellman (1995), the effects of retrieval practice of an item in one category (e.g., GREEN—emerald) were examined on retrieval of an item paired with a different category cue at encoding (e.g., SOUP—mushroom). In Anderson et al. (2000), the effects of practice of an item from a category (e.g., RED—brick) were tested for other items from the same category (e.g. tomato), but with the independent category cue FOOD used at the final test stage. We discuss each methodology in turn.

In Anderson and Spellman's (1995) paradigm, participants studied lists of category-exemplar pairs in the manner introduced by Anderson et al. (1994). However, within the items studied there were systematic relations between the items. So, in Experiment 2, for example, participants studied 6 items associated with the cue GREEN and 6 with SOUP. Critically, half of the GREEN items were vegetables, as were half of the SOUP items. Thus, there was an implicit category of VEGETABLES within the experiment, although the cue VEGETABLES itself was never presented. In the retrieval practice phase, participants practiced retrieval for the non-vegetable items from one of the presented categories (e.g., GREEN—emerald). This results in four kinds of items at test. As in Anderson et al. (1994) there were practiced items (RP+), and unpracticed items from the same category (RP−). However, whereas Anderson et al. (1994) merely had unpracticed items from unpracticed categories (NRp items), Anderson and Spellman's (1995) methodology led to NRp items that are similar to the RP− items (designated NRpS), and NRp items that are dissimilar to the RP− items (NRpD items).

The results indicated that retrieval practice of GREEN—emerald led to retrieval-induced forgetting for the other items associated with that category, which is the standard retrieval-induced forgetting effect seen for the RP− items (e.g., GREEN—lettuce). In addition, forgetting was also seen for the items from the other category that were similar to the RP− items (i.e., NRpS items such as mushroom), even when cued with SOUP. Thus, since forgetting was seen for the NRpS item using a non-practiced category cue, Anderson and Spellman (1995) argued that the forgetting effect must be due to inhibitory processes acting upon the memory representations, rather than due to an associative process that links the category and the item. They proposed a feature competition explanation of this pattern, in which the cue (GREEN) elicits the desired targets (e.g., emerald), whilst at the same time inhibiting those features of the undesired items (e.g., lettuce) that are not shared with the practiced items. All three non-retrieved items are vegetables, and so the vegetable qualities of these unwanted green items are inhibited. These inhibited vegetable-features will be shared with other vegetables, such as mushroom, and so the result is that mushroom is also inhibited. Thus mushroom is difficult to recall, even when cued with SOUP.

Whilst this experimental methodology is elegant, and the argument persuasive, we reserve judgment about the inhibitory account of retrieval-induced forgetting observed in this study. One reason is that the results are perhaps surprising given the previous findings reported by Anderson et al. (1994). They demonstrated that weak category exemplars (e.g., FRUIT—nectarine) were not inhibited by practice of other category members, whether they are strong (e.g., orange), or weak (e.g., fig) members of the same category. Thus according to Anderson et al. (1994) practice of orange does not inhibit concepts as similar as nectarine or guava. However, in Anderson and Spellman (1995) practice of dollar, emerald and lawn from the cue GREEN produces forgetting for artichoke, lettuce, and pepper from the same cue, which in turn produces forgetting of mushroom, onion, and tomato from the cue SOUP. Although norms are not available, it does not appear likely that artichoke, lettuce, and pepper are strong members of the category GREEN. Thus, following Anderson et al. (1994) they would not be expected to compete for retrieval from the cue GREEN, and so should not be inhibited. Thus, in the context of the Anderson et al. (1994) findings, the inhibition reported in Anderson and Spellman (1995) appears unexpected given the materials used.

A second concern arises from reconsideration of Anderson and Spellman's (1995) data across three crucial experiments. Experiments 2 and 4 in their paper demonstrated second-order inhibitory effects along the lines discussed above. In each case, the comparison was made between conditions in which the critical NRpS (mushroom) item was studied either in the context of related or unrelated RP− items. In each case, recall was lower in the presence of similar RP− items, thus supporting Anderson and Spellman's claims of inhibition being spread because of similarity between RP− and NRpS items. The key difference between the two experiments was that Experiment 4 controlled for output interference effects, which the authors acknowledge was a potential confounding explanation for the data reported for Experiment 2.

Table 1 reproduces the data from Anderson and Spellman (1995), taken from Table 2, Table 4 of their paper. As well as providing the critical comparisons that Anderson and Spellman focused on, this table also contains data from Experiment 3a, which was conducted without the key retrieval practice phase. This experiment enabled the authors to demonstrate that the previous forgetting effects seen in related conditions were not because of some aspect of list structure across related and unrelated lists. However, this condition also provides a useful baseline against which the previous effects on Rp− and NRpS items can be interpreted, since in that experiment there was no retrieval practice, and so no retrieval-induced forgetting should occur. This baseline condition produced a recall rate of 48% for both related and unrelated conditions.

The crucial data concern the recall rates for NRpS items, since they were used by Anderson and Spellman (1995) to argue for an inhibitory account of retrieval-induced forgetting. These data are boxed in Table 1, although performance on the other critical items is included for completeness. First, consider the data from the related conditions, in the upper box, which are the conditions in which retrieval-induced forgetting is predicted. Compared to the baseline recall rate, retrieval practice produced an apparent forgetting rate of 11% in Experiment 2, where output interference was not controlled, but only 3% where output interference was controlled in Experiment 4. Now, turn to the unrelated conditions, where no retrieval-induced forgetting is predicted, and which were used by Anderson and Spellman as the contrast conditions in Experiments 2 and 4. Since no retrieval-induced forgetting is predicted, one might expect similar performance to the baseline level of 48% observed when no retrieval practice occurred. However, recall of NRpS items was 4% above baseline for Experiment 2, and 9% above baseline in Experiment 4. Thus, the robust retrieval-induced forgetting effects reported by Anderson and Spellman for Experiments 2 and 4, obtained by contrasting the related and unrelated conditions in each experiment, seem to be due to an unexpectedly high level of recall for the NRpS items in the unrelated condition, rather than due to suppression in the related condition. Whilst it is beyond the scope of the present discussion to fully explain this pattern, it is clear that the majority of the effect reported by Anderson and Spellman (1995) is not consistent with an inhibitory effect. Rather, it appears to be consistent with an inexplicable retrieval benefit afforded to NRpS items in the unrelated condition. Unfortunately, the other studies that have demonstrated second order inhibitory effects (Anderson & Bell, 2001) failed to include a no-practice control, and so it is impossible to know whether or not a similar effect could explain those data.

A final concern with this second-order inhibitory effect is that there is one published failure to replicate this effect, despite using a design with greater numbers of items per category, and more participants in the study than in the original Anderson and Spellman (1995) paper (Williams & Zacks, 2001). Whilst this failure to replicate is only an isolated finding, together with the concerns raised above, it seems reasonable to maintain a reasonable level of doubt about the reliability of the second-order inhibitory effect. Since this effect is crucial in establishing inhibition as the mechanism underlying retrieval-induced forgetting, it is clear that further evidence is required.

Although the Anderson et al. (2000) paradigm introduces an entirely new independent probe at the final recall test (i.e., one not presented during study), similar problems may complicate interpretation of these findings. In this paradigm, participants study and practice in the standard manner. However, in the final test condition using an independent probe, participants are tested using semantically related cues that were not present at encoding or practice. So, for example, RED—brick and RED—tomato are studied. Following retrieval practice of RED—brick, tomato is less likely to be recalled, either from the cue RED, or the cue FOOD, which was never present at study. In this manner, cue-independent forgetting is demonstrated, and hence inhibition. However, as in Anderson and Spellman (1995), the category RED was studied with multiple (eight) items, half of which were foods. No food items appeared in any other category. Thus food is a highly salient aspect of the RED category, and participants are likely to have noticed this, and to have formed an association between RED and FOOD. Now imagine a participant faced with recalling items from the cue FOOD. If their retrieval strategy is to remember that foods were uniquely associated with RED, then they may use RED as a proxy retrieval cue rather than using FOOD directly. Why should they do this? One reason is that this independent probe technique is effectively a version of an encoding specificity experiment (Thomson & Tulving, 1970; Tulving & Thomson, 1973) in that participants are given a strong associate of an item as a memory cue for an item studied with another cue. However, as Tulving and Thomson (1973) showed, this is a poor cue, relative to the originally encoded cue. Thus, one might expect the cue FOOD not to be a very useful retrieval cue to participants. Given that, what can participants do, other than to try to use a more effective cue? The most recent relevant aspect of the experiment will have been the retrieval practice phase, which means that the category cue RED is highly accessible. Given that this was the cue originally paired with the forgotten item, and that half the studied red things were food, it is possible that at least some participants try to use the cue RED when presented with the cue FOOD. The consequence is that this merely reinstates the originally practiced retrieval cue, and so the non-practiced items are likely to be poorly recalled. Thus an associative account of the supposedly independently cued forgetting effect is possible, based on covert-cueing.

Anderson and his colleagues (Anderson & Bell, 2001; Anderson et al., 2000; Anderson & Spellman, 1995) have argued that such covert-cueing effects cannot explain their data for two reasons. One is that those who report using such a strategy show numerically, but not statistically, less inhibition (Anderson, Bjork, & Bjork, 2000). However, this argument rests upon self-reported use of a covert-cueing strategy in which participants report on a rating scale how often they mentally scanned through previously encountered category names to find the sought after item. However, this can only be completed after recall, and so it is entirely plausible that the reported strategy is driven by the success or failure at recall, rather than vice-versa. That is, those who perform poorly (i.e., show forgetting effects) deny the use of the strategy, whilst those who do well (i.e., show less forgetting) claim that they did use the strategy. Thus, there is good reason to question the usefulness of the self-report data. The second argument is that covert-cueing predicts forgetting where none is found (Anderson & Bell, 2001). For example, participants faced with the non-practiced category SOUP may recall the overlap with the previous category GREEN, via the shared vegetable items. This would then lead them to recall the practiced GREEN exemplars, so blocking recall of both SOUP exemplars. Crucially, this argument applies equally to items that share features (e.g., mushroom) as well as those that do not (e.g., turkey). Thus forgetting for NRpD items is predicted, but was not observed in either Anderson and Spellman (1995) or Anderson and Bell (2001). However, as we noted above, Anderson and Spellman's (1995) data are not as clear cut as they might be, when no-practice baseline is taken into account, and Anderson and Bell (2001) did not report any no-practice baseline data.

Thus, whilst it is not clear that a covert-cueing account can fully explain the pattern of data in the literature, it is also the case that the inhibitory account is open to challenge. Both independent probe methodologies can be challenged with regards how well they provide an independent cue for the putatively inhibited item. Their use of the term independent refers to the non-occurrence of the critical cue during the first two stages of the experimental paradigm, which thereby produces a cue that is independent of any associations created in the study or practice phases. However, there is a stronger sense of the word independence, which is logical or statistical independence. Two events that are logically independent are completely unrelated, in the sense that the outcome of one event does not influence the outcome of the other. On this criterion the previous methodology fails, since it is clear that the independent probes are not unrelated to the previously experienced events. For example, someone given the cue word VEGETABLE is more likely to go on to generate the item onion than they are siren, whatever their memory for those two items. Similarly, whatever their memory they are more likely to think of the previous category cue GREEN than the previous category cue LOUD. This dependence between the independent cues and the list structure is an inevitable consequence of relying upon prior semantic associations between independent cues and to-be-remembered items, and associations between the items themselves.

In three experiments presented here, we adopt a different methodology in which to-be-remembered items are associated with an item that is independent in all respects, save for their prior episodic co-occurrence. In Experiments 1 and 2, participants are presented with category—exemplar pairs as in the standard retrieval-induced forgetting paradigm. However, in addition, each pair is presented along with an independent item, namely a photograph of a face. Each face is arbitrarily paired with a category—exemplar pair, and so is formally independent of the to-be-remembered information. That is, there is no information in the face that will cue the particular exemplar beyond the episodic relationship established by their co-occurrence. There is no hidden structure to the categories. Thus, these cues are statistically and semantically independent of the both categorical cues, and all target items. In Experiment 1, the retrieval practice phase replicates that used by Anderson et al. (1994), with retrieval of category exemplars cued by categories and word stems alone. No faces are presented at practice. This means that at the final test stage it is possible to cue memory for the items either by category cue (as in the standard retrieval-induced forgetting paradigm), by the face originally paired with the exemplar, or by a joint category plus face cue. In Experiment 3, we used words with no prior association with either the exemplars or the category cues as independent episodic cues in the same manner. Additionally we removed any association between the episodic cue and the category cue by presenting the episodic cue together with the exemplar words in a separate list, so that there was never any co-occurrence of the episodic cues and the category cues.

The predictions are relatively straightforward for the inhibitory account in each case. Since the retrieval practice phase involves category cueing, competition will be invoked and the non-retrieved items will be inhibited. Consequently, poorer recall should be demonstrated for those RP− items, whatever the retrieval cue, even if that previous cue is one that was only ever episodically associated with the forgotten item. On the other hand, if retrieval-induced forgetting is due to an associative bias caused by retrieval practice, one would expect a different pattern. As before, the use of category cues at test should produce retrieval-induced forgetting, but now face cues should not demonstrate retrieval-induced forgetting effects. The prediction regarding combined cues is less clear cut. If participants rely heavily on the category cues, ignoring the faces at test, then retrieval-induced forgetting may occur. On the other hand, the extent to which the faces are useful cues should be revealed by the reduction in the retrieval-induced forgetting effect relative to the category cue condition. That is, participants may suffer a retrieval-induced forgetting effect due to the categories, but be able to overcome it by use of the faces.