Introduction

Making an analogy involves recognizing a common relational system between two different situations (Gentner, 1983; Gentner & Markman, 1997; Holyoak, 1984; Holyoak & Thagard, 1995). Typically, a more familiar episode (the base analog) is employed to enhance our comprehension of a less understood episode (the target analog). For example, a situation in which a colleague of ours was eventually not offered a professorial position at an important university and then claimed that this position was not in her preferred area, could bring to mind Aesop’s fable of the fox and the unreachable “sour” grapes. These two situations could be considered analogous to the extent that they are cases in which the protagonist tries to achieve a goal, fails, and then retroactively judges that the goal had not really been desirable after all. Shared relational patterns like these constitute the basis for analogical relatedness and are referred to as structural similarities. On the basis of these shared systems of relations, an alignment can be established between the objects playing parallel roles in their respective situations (e.g., the fox and our colleague or the grapes and the new position). As this example illustrates, the corresponding objects in the base and the target need not resemble one other. A tale will not need to involve animals and fruit in order to be analogous to the sour grapes story. What is important is that base and target entities hold like roles in the matching systems of relations. Since similarities at the level of objects do not constitute a requirement for analogical relatedness, they are referred to as surface similarities.

Analogical retrieval refers to the phenomenon through which, while thinking about a topic or scenario, people are reminded of a structurally similar piece of knowledge or past experience. Intelligent access to structurally similar situations in long-term memory (LTM) constitutes a central component of knowledge transfer and creativity (Gentner, Holyoak, & Kokinov; 2001; Hofstadter & Sander, 2013; Loewenstein, 2010). The sad conclusion from a large body of laboratory research is that, in contrast to the ease with which superficially similar analogs (or near analogs) come to mind, the retrieval of superficially dissimilar analogs (or distant analogs) often fails (Catrambone, 2002; Gentner et al., 1993; Gick & Holyoak, 1980, 1983; Holyoak & Koh, 1987; Ross, 1984, 1987; Wharton et al., 1996). The fact that people can easily understand analogies between structurally similar episodes maintaining no superficial similarities led to the conclusion that the retrieval stage constitutes the main bottleneck in analogical thinking (Gentner et al., 1993; Gick & Holyoak, 1980).

With the turn of the century, however, naturalistic studies have started to obtain unprecedented proportions of distant analogizing, leading their authors to reinterpret the scarcity of distant retrievals in laboratory studies as resulting from extremely adverse experimental conditions (see, e.g., Dunbar, 2001; Hofstadter & Sander, 2013). According to this perspective, under conditions more representative of those taking place in everyday life, the gap between near and distant retrieval should prove to be narrower than in traditional experimental studies. The present study was carried out to assess whether the encoding of base and target analogs in terms of the same relational category – a frequent circumstance in everyday life – promotes better access to superficially dissimilar analogs.

As in the study of memory at large, the experimental procedure most widely used to investigate analogical retrieval comprises two phases. During the learning (or encoding) phase, participants receive one or several base analogs interleaved among distractor situations. During the retrieval phase, participants receive the target analog, and the experimenters assess whether its processing elicits the retrieval of the critical situation learned during the previous phase. While studies interested in spontaneous analogical retrieval do not suggest to participants the usefulness of recalling analogous situations, studies interested in our ability to voluntarily search for analogous situations prompt the retrieval of analogous cases.

Studies of spontaneous analogical retrieval often employed Gick and Holyoak’s (1980) Military story and Duncker’s (1945) Radiation problem as base and target analogs, respectively. The military story told about a dictator who governed a country from a fortress, which could only be captured with a large number of soldiers. A rebel General had a large enough army, but he learned that the radial roads leading to the fortress were mined. Hence, he could not send the whole army through a single road. After dividing his army into smaller groups, the General placed them on different roads and had them converge on the fortress at the same time, thus avoiding the detonation of the mines. The radiation problem used as target analog told about a patient who had an inoperable tumor in his stomach. Even though the patient’s surgeon had a type of ray that could destroy the tumor if applied at a very high intensity, rays of such intensity would also destroy surrounding tissues, which needed to be preserved. Participants had to envision ways of using the rays to destroy the tumor, but without sacrificing the surrounding tissues. Gick and Holyoak (1980, 1983) found that only a small proportion of participants (around 30%) applied a convergent strategy to the target problem. Studies enforcing a stronger contextual separation between the encoding and the transfer phases obtained even lower transfer rates (just over 10%; e.g., Keane, 1987; Spencer & Weisberg, 1986). Studies comparing the retrieval of near versus distant analogs also showed that near analogs (e.g., a base story nearly identical to the tumor problem but in which the tumor was located in the brain instead of the stomach; Keane, 1987; see also Holyoak & Koh, 1987) were much more likely to be spontaneously retrieved (above 80%).

Studies interested in voluntary reminding have tended to work with simple stories instead of problems. In this story-reminding tradition, participants read a large set of stories within which the critical base analogs are buried among several distracter items. Upon receiving a second set of stories, participants are asked to indicate, for each new story, any stories of the previously presented pool of which they are reminded. In their most widely known set of materials, Gentner et al. (1993) included as a base analog a tale in which an old hawk was attacked by a hunter who needed feathers for his arrows. After the hunter missed, the hawk glided down and offered to give him a few feathers,  for what the hunter pledged never to shoot the hawk again. The near analog included minor variations such us an eagle instead of a hawk, a sportsman instead of a hunter, and a crossbow instead of a bow. In the distant analog, a warlike country wanted its neighbor’s new computers and attacked it with missiles. After the aggressor failed, the attacked country offered to sell some of its computers to it. Very pleased, this country promised never to attack again. Across experiments, Gentner et al. (1993) obtained retrieval rates of nearly 60% for close analogs, versus 20% for distant analogs. With few exceptions (see, e.g., Raynal et al., 2020), the difficulty of retrieving distant analogs obtained in problem-solving studies is replicated even when people are prompted to concentrate on a set of base items that, albeit large enough to impede a serial type of search, represents only a minimal fraction of LTM (Catrambone, 2002; Wharton et al., 1996). Taken collectively, the results of the problem-solving and the story-reminding experimental traditions were taken as evidence that analogical retrieval rarely occurs with a lack of surface similarity.

Despite the robustness of traditional laboratory results, it could be argued that the disappointing levels of distant retrievals might have originated in unfairly adverse experimental conditions. One disadvantageous feature of the materials used in the experimental tradition is that they do not tend to favor the detection and articulation of the abstract generic schemas that could facilitate purely structural remindings. When salient aspects of the base analog are similar to salient features of the target, there will be many potential cues to retrieve the former during the processing of the latter. For example, the radiation problem is likely to call to mind prior knowledge about related medical problems involving tumors, rays, physicians, patients, and surgeries. In contrast, distant analogs stored in LTM (e.g., the military problem) will lack any resemblance at the level of such shallow characteristics, keeping its similarities at a more conceptually deep, not immediately evident level. If some of these structural features of base analogs happen to be abstracted during their initial encoding, these analogs could become more likely to be cued by a superficially dissimilar target. Following on with the military story, its abstract schema would include the need to use an intense force to overcome a central target, plus the fact that employing the force at such intensity would endanger elements that need to be preserved. In case the subsequent presentation of the target elicited a similar schema, the distant analog could potentially be accessed by virtue of having been originally indexed under the described schema (the “learning to encode principle,” Medin & Ross, 1989). One of the limitations of the situations employed in the experimental tradition is that they seldom elicit this kind of abstraction.

But the difficulties faced by participants of the experimental tradition reach even further. In those infrequent cases where participants engage in deriving abstract representations, these conceptualizations will only serendipitously match the exact schema that will later be elicited by the target (Dietrich, 2000). Using the military story as an example, encoding the goal of capturing the fortress as a case of “overpowering an object by a force” would certainly match the target goal of “destroying a tumor with rays” (Gick & Holyoak, 1983). But things would be different if instead of the tumor problem, participants were tasked with figuring out how to bring a desired flow of water to a town while attending to the restriction that sending such a flow of water through either of the available canals alone would flood the surrounding lands. In order to retrieve the military story during the processing of this particular problem, it would have been better to conceptualize the goal of capturing the fortress as a case of, say, “carrying a resource to a destination.” Hence, a second reason behind the scarcity of distant retrievals in experimental studies may be that the schemas that participants construct “on the fly” during the processing of the base analogs are unlikely to match the schemas that will later be elicited by the target.

In contrast to the rather uninteresting character of the stories employed in most experimental studies, it has been argued that the situations on which we base our daily-life analogies tend to be familiar and personally meaningful (Hofstadter & Sander, 2013). This increased meaningfulness might conceivably aid in pushing surface information to the background, and bringing more profound structure to the fore. Even in those cases where an immediate generalization either failed to occur or turned inconveniently idiosyncratic, the fact that personally significant episodes tend to be accessed over and over again is likely to drive generalization in conservative ways that encompass both the original episode and those later situations that prompted its retrieval (Ross & Kennedy, 1990). Hence, a third negative aspect of the materials employed in the laboratory (e.g., the military story or the story of Karla the hawk) could be their artificial character.

Even though the conditions and materials confronted by participants in traditional laboratory experiments could be considered suboptimal for eliciting the particular representations that will be needed for future analogical retrieval, these difficulties could disappear in those cases where naturally acquired base and target analogs automatically elicit the activation of the same well-established schema, which could serve to clearly demarcate common structural features of the events under consideration. As an example, consider a base analog describing a situation in which a man makes elaborate preparations to spend the inheritance of a still-living relative, before confirming that his relative included the man in her will. This situation might be likened to the saying “don’t count your chickens before they are hatched,” therefore facilitating the encoding of this situation in terms of the abstract schema underling that saying (i.e., you shouldn’t get your hopes up or make plans based only on assumptions, because it can lead to disappointment). If a subsequent situation features a person preparing a graduation party before receiving the grade of her last exam, there is also some probability that such a story would activate the referred saying and schema. The application of the same schema to both situations would ensure a similar abstract encoding, wherein superficially mismatching features are de-emphasized. When these conditions are met, it could be expected that the retrieval of distant analogs would be facilitated, as compared to conditions wherein appropriate schemas are not available.Footnote 1

For experimental materials such as Gick and Holyoak’s (1980) military story, participants do not possess abstract schemas such as “applying a resource at maximum strength can capture a central target, but with the consequence of destroying surrounding elements.” In contrast with these rather unique situations, real-life events are often instances of overlearned schemas that are robustly established in LTM, and which are activated almost obligatorily during the processing of each of these instances. For example, schema-governed categories (SGCs; Gentner & Kurtz, 2005; Markman & Stilwell, 2001) are relational categories that specify the structure of situations or events in terms of a network of semantic interdependencies that hold among the constituents of the concept in question. In contrast to members of entity categories, which share a set of probabilistic features and feature correlations, members of SGCs such as assassination share a structure that can be instantiated by exemplars that are superficially very different (Gentner & Kurtz, 2005; Goldwater et al., 2011; Markman & Stilwell, 2001), such as “Fred thrust a knife into Gina’s heart,” “Mary had Bob drink poison,” or “The offender disconnected the patient’s oxygen supply.” As in the case of proverbs, overlearned schemas of this kind de-emphasize surface content in favor of the relational structure that renders an event and instance of the category, therefore promoting a uniform encoding of relations (Jamrozik & Gentner, 2020; Kurtz & Honke, 2020). This effect could be especially robust when the category is lexicalized and its name is activated during the act of categorization. From this ability of lexicalized relational categories to bring structural features to the fore, it follows that distant exemplars of SGCs should be more easily recalled from LTM than the base analogs typically employed in psychological experiments. For a long time it has been argued that the advantage of experts over novices in distant retrieval results from the fact that experts tend to represent domain phenomena in terms of domain principles, thus promoting a uniform relational encoding (e.g., Forbus et al., 1995), and recent evidence attests to this possibility (Goldwater et al., 2021). It is reasonable to expect that when people encode familiar situations with expert-like schemas, distant retrieval would be easier than in the typical studies of the experimental tradition.

The minor attention received by relational categories in the literature about concepts might suggest that they are not abundant in everyday life, but in fact the opposite is true. For example, informal ratings of the 100 highest-frequency nouns in the British National Corpus revealed that about half are relational nouns, with SGCs making up an important portion of them (Asmuth & Gentner, 2005). A second and related question refers to how frequently people draw analogies between exemplars of SGCs. Even though the relative frequency of this kind of analogy has not been established, a quick look at any newspaper will reveal just how frequently they appear: at the time of this writing, one can find numerous discussions on whether the quarantine implemented in Argentina was analogous to the one implemented in Uruguay, if the debt contracted by Mauricio Macri is analogous to the one incurred by Cristina Kirchner, or if the birthday party celebrated by the president during the lock-down represents a violation of the quarantine restrictions analogous to the organization of a popular demonstration against the government, carried out by the Opposition. In the domestic arena, one can easily find oneself discussing issues such as whether a current marital crisis is analogous to one that occurred in the past, whether the action slip of pouring hair gel on the toothbrush is analogous to that of placing the car keys in the refrigerator, or whether children’s addiction on their gaming consoles is analogous to adults’ commitment to work (see Jamrozik and Gentner (2020), Oberholzer et al. (2018), and Raynal et al. (2018) for a shared opinion about the profusion of these types of analogies in daily life). As SGCs can sometimes be well captured by a single lexical item, it would seem that comparisons between exemplars of SGCs involve a mapping between single relations. Hence, it could be argued that they could hardly be regarded as analogies, on the grounds that true analogies involve systems of relations and not just single relations. However, an analysis of the meaning of SGCs often reveals complex systems of relations. As an example, the SGC robbery tends to involve an intentional agent whose action results in something that belongs to the patient becoming the property of the agent, but without the consent of the patient.

As dictated by their subject of inquiry, studies assessing the retrieval of naturally acquired exemplars of schema-governed categories must resort to free-production paradigms wherein participants are presented with a target analog, with the task of recalling related episodes from their general, extra-experimental memory.Footnote 2 In an early study on the retrieval of SGC exemplars, Olguín, Tavernini, et al. (2017a) presented participants with an exemplar of a SGC (e.g., “I painted a portrait for my girlfriend,” an instance of the SGC expression of love) and asked them to come up with “me too” analogies (see Hofstadter & Sander, 2013) by way of recalling analogous autobiographical episodes. While 56% of the analogical responses were semantically distant from the target (e.g., “I prepared my girlfriend’s favourite breakfast”), 44% were semantically close to the target (e.g., “I wrote a poem for my girlfriend”), a result that provided preliminary evidence for the advantage of SGC exemplars in enabling superficially unconstrained analogical retrieval.

In a very similar study, Raynal et al. (2018) presented participants with stories like the following: “I had the idea to answer that I had forgotten my glasses when the photographer offered that I could go to his exhibition. The reality is that I had my glasses with me but I did not want to go there” (an instance of excuse). Participants had to report autobiographical episodes that these situations reminded them of. Almost 50% of the responses were distant analogs, whereas 12% of the responses were near analogs. Even more compellingly than in Olguín et al. (2017a), their results suggest that when the base and the target have been encoded in terms of a shared SGC, surface similarity no longer represents a hard precondition for analogical retrieval.

Despite the merit of the studies by Olguín et al. (2017a) and Raynal et al. (2018) in addressing the retrieval of daily-life exemplars of SGCs, their method is not ideally suited for assessing the role of surface similarity during retrieval, since the obtained pattern may merely be the result of having more distant than near analogs in memory. How many instances of producing an artwork in the service of expressing love could Olguín et al.’s participants have available in memory? Most likely, fewer than the number of expressions of love coming from all other domains (e.g., holding the lover’s hand, wearing her favorite clothes, or hiring a band of Mariachis). Likewise, how many instances of having invoked visual difficulties for not attending a cultural activity could Raynal et al.’s (2018) participants have available in memory? Most probably fewer than the sum of excuses coming from any other domain (e.g., claiming being stressed for not initiating a diet, not feeling well for not attending a class, being urged not to keep the house in order, etc.). Assessing the extent to which surface similarity determines the retrieval of naturally acquired base analogs would require knowing not only the number of far and near sources that were retrieved, but also the number of instances of both types of base analogs that were potentially available for retrieval (Trench & Minervino, 2015). Even though this is factually unfeasible, an equivalent way of comparing the retrieval likelihood of close versus distant analogs could consist of documenting the availability of either a close or a distant source analog in each participant’s LTM, and then comparing the retrieval probabilities of these particular episodes during the processing of an analogous target.

In order to reassess whether an encoding of daily-life autobiographical episodes in terms of a shared SGC reduces the weight of surface similarities during retrieval, we adapted the hybrid paradigm developed by Trench and Minervino (2015, Experiment 2) so as to form two groups of participants who, as determined by a questionnaire, had experienced an instance of a particular SGC (e.g., robbery) from one of two domains of knowledge (e.g., vehicles or informatics). During a temporally and contextually separated session, the experimenters presented participants with a further exemplar of the SCG that participants had experienced. While the SGC to which the target pertained was in both conditions the same as the one reported in the first session (e.g., robbery), the domain of the target only matched the domain of the episodes reported by participants of one of the groups (e.g., vehicles). Experiment 2 had the same design and procedure as Experiment 1, with the difference that base and target analogs were composed of two facts pertaining to two different SGCs and connected by a causal relationship (e.g., effort causes achievement).

Experiment 1

Method

Participants and design

One hundred and thirty-nine students of Psychology (mean age = 21.8 years, SD = 2.92) volunteered to participate in the experiment. They were randomly assigned to either the intradomain condition or the interdomain condition. The final sample was composed of 48 participants of the intradomain condition and 48 participants of the interdomain condition who reported having experienced an episode of a particular SGC. The dependent variable was whether the target analog received during the second phase elicited the retrieval of the situation that had been reported during the first phase.

Materials

Four sets of materials were built, each one comprising a target situation and two cues (see Table 1). The target analog of each set described a simple situation in which a character had experienced a situation pertaining to a SGC and a specific domain (e.g., “I have a friend who has experienced an episode of unpunctuality: He arrived late to take a plane”). The two cues of the questionnaire presented during the first phase of the study were intended to check the availability of naturally acquired autobiographical episodes belonging to the same SGC as the target. While near cues aimed at identifying situations belonging to the same domain as that of the target (e.g., having arrived late to take a transport), distant cues aimed at identifying instances of the same SGC but belonging to a domain that was different from that of the target (e.g., having arrived late to a doctor’s appointment). For the first phase, two alternative two-page booklets were compiled: the intradomain version, placing the semantically near cues of the four sets, and the interdomain version, collecting the semantically distant cues of the same four sets. For the second phase, a single sheet included the target situation of one of the sets plus two filler tasks.

Table 1 Cues and target situations, Experiment 1
Full size table

Procedure

After providing signed informed consent, participants carried out the first phase of the experiment. This phase was presented to participants as a study on memory of autobiographical episodes, and was administered during an introductory class on Cognitive Psychology. After each of the four cues, written instructions requested participants to write down an autobiographical episode that corresponded to the cue, in case that they had experienced one. They had to provide as many details as possible about what exactly occurred, when it happened, and where it took place. Participants had 15 min to complete the task. Two independent judges analyzed each of the answers given by participants to determine the SGC to which the reported event pertained (e.g., unpunctuality). Judges received each response along with the four SGCs that were used for constructing the sets of materials plus a “none of the above” option. They had to indicate to which of the four SGCs each episode belonged, in case they considered that it did. Responses in which at least one of the judges failed to categorize a response as an instance of the critical SGC (8%) were dropped from further analysis. For those that were not discarded, judges had to determine if the situation reported pertained to a particular domain (e.g., transport or health advice). The list of domains that judges received comprised the eight domains that appeared within the four sets of materials, plus a “none of the above” option. Responses in which at least one of the judges failed to categorize a response as an instance of the critical domain (6%) were dropped from further analysis. Finally, participants of each condition were semi-randomly subdivided until obtaining 15 participants who reported a base analog for Set 1, 15 participants who reported a base analog for Set 2, 15 participants who reported a base analog for Set 3, and 15 participants who reported a base analog for Set 4 (Fig. 1 shows the flow of participants throughout the experiment).

Fig. 1
figure 1

Flow of participants through Experiment 1

Full size image

In order to enforce a contextual separation between the first and the second phases of the study, an experimenter that was not present during the first phase sent an e-mail invitation to participants of the above groups to take part in a study about story-telling abilities. The second phase took place at the Cognitive Studies Laboratory of the University. After receiving a brief presentation about the story-telling study, each participant received two filler tasks: one of choosing between two endings for a story and one of selecting a character for a hypothetical story. Upon completing these tasks, each participant received the target analog that corresponded to the set of materials to which that particular participant was assigned. Participants were asked to report one analogous story that had happened to them, if they had had one, trying to provide a complete description of it. They were allotted 15 min to complete the second phase. Once they had finished, they were asked to report if, prior to receiving the last situation, they had noticed a connection between the present story-telling study and the memory task performed during the previous week. Those who answered positively were withdrawn from the analysis. This way, we ensured that the eventual retrieval of the critical base analog was based on general search mechanisms and not on an episodic association between the two phases.Footnote 3

For each condition, we formed a group of 12 participants who had reported a base analog for Set 1 during the first phase, a group of 12 participants who reported a base analog for Set 2, a group of 12 participants who reported a base analog for Set 3, and a group of 12 participants who reported a base analog for Set 4 (see Fig. 1).

Data analysis

Two judges received the transcripts of participants’ episodes reported in the first phase coupled with the episode reported by the same participant during the second phase. Judges had to decide whether or not the two episodes were the same. They were asked to pay special attention to whether there was a coincidence between the details provided by participants for each event. Judges agreed in 94.7% of the cases, resolving cases of disagreement by discussion.

Results and discussion

Base analogs that were semantically close to the target were retrieved in 72.91% of the cases, while semantically distant sources were retrieved in only 12.5% of the cases, χ2(1, N = 96) = 35.803, p < .001, φ = .61. The retrieval rates of near and distant exemplars of the SGC that applied to the target were very similar to the ones obtained in traditional studies of analogical problem-solving (well above 80% vs. just above 10%) and also similar to the ones obtained in laboratory studies of story-reminding (around 60% vs. about 20%). Table 2 shows the retrieval rates of intra- and interdomain source analogs for each of the sets of materials.

Table 2 Retrieval rates by set, Experiment 1
Full size table

In Experiment 1 we dealt with analogies between two exemplars of an SGC. However, it is often the case that the relational systems that give rise to analogical inferences comprise pairs of causally related SGCs such as misconduct-sanction, investment-payoff, effort-success, temptation-sin, economic crisis-recession, achievement-celebration, illness-treatment, error-justification, etc. The objective of Experiment 2 was thus to assess whether the results of Experiment 1 generalize to analogies between relatively more complex situations, in which an exemplar of a SGC is causally connected to an exemplar of a different SGC.

Experiment 2

Method

Participants and design

One hundred and fifty-two students of Psychology (mean age = 22.7 years; SD = 3.61) volunteered to participate in the experiment. They were randomly assigned to the intradomain and interdomain conditions. The final sample consisted of 44 participants of each condition who reported having experienced an instance of a particular system of SGCs. The dependent variable was whether the target analog received during the second phase elicited the retrieval of the situation that had been reported during the first phase.

Materials

Four sets of materials were built, each one comprising a target situation and two cues (see Table 3). The target analogs of each set were situations in which an instance of an SGC (e.g., bad behavior) led to an instance of a different SGC (e.g., punishment) in the same domain (e.g., school). The two cues were intended to check the availability of naturally acquired autobiographical episodes composed of instances of the same SCGs as those of the target, and connected by the same causal link. As in Experiment 1, near cues aimed at identifying situations taking place within the same domain as that of the target (e.g., school), whereas distant cues aimed at identifying situations taking place within a domain that was different from that of the target (e.g., home).

Table 3 Cues and target situations, Experiment 2
Full size table

Procedure

The two phases of the experiment followed the same procedure as in Experiment 1. Two independent judges analyzed participants’ responses to the availability questionnaire in order to determine whether or not they pertained to the SGCs included in the cue. Judges received each response along with the four systems of SGCs that were used for constructing the sets of materials plus a “none of the above” option. They had to indicate to which of the four systems of SGCs each event belonged, if they considered that it did. Responses in which at least one of the judges failed to categorize a response as an instance of the critical system of SGCs (5%) were dropped from further analysis. For those that were not discarded, judges had to determine if both the cause and the effect of the situation reported pertained to one particular domain (e.g., school). The list of domains that judges received comprised the eight domains that appeared within the four sets of materials, plus a “none of the above” option. Responses in which at least one of the judges failed to categorize a response as an instance of the critical domain (9%) were dropped from further analysis. Participants of each condition were semi-randomly subdivided until obtaining 15 participants who reported a base analog for Set 1, 15 participants who reported a base analog for Set 2, 15 participants who reported a base analog for Set 3, and 15 participants who reported a base analog for Set 4. As in Experiment 1, participants who admitted having noticed a connection between the current phase and that of the previous week prior to receiving the target situation were withdrawn from the analysis. For each condition, we formed a group of 11 participants who reported a base analog for Set 1, a group of 11 participants who reported a base analog for Set 2, a group of 11 participants who reported a base analog for Set 3, and a group of 11 participants who reported a base analog for Set 4. Figure 2 shows the flow of participants through the experiment.

Fig. 2
figure 2

Flow of participants through Experiment 2

Full size image

Data analysis

Two judges received the transcriptions of each participant’s episode reported in the first phase coupled with the episode reported by the same participant during the second phase, and had to decide whether or not the two descriptions referred to the same event. They were asked to pay special attention to whether there was a coincidence between the details provided by participants for each event. Judges agreed in 91.5% of the cases, resolving cases of disagreement by discussion.

Results and discussion

Near analogs were retrieved in 27.27% of the cases during the second phase, while distant analogs were retrieved in only 4.55% of the cases, χ2 (1, N = 88) = 8.494, p < .01. φ = .31. Results replicate those of Experiment 1 with materials that included systems of SGCs instead of isolated SGCs. Therefore, they provide further evidence that having encoded an autobiographical base episode and a target analog in terms of the same lexicalized SGCs neither improves distant reminding nor diminishes the effect of surface similarities on analogical retrieval. Table 4 shows the retrieval rates of intra- and interdomain source analogs for each of the sets of materials.

Table 4 Retrieval rates by set, Experiment 2
Full size table

Although the ratio of near and distant analogs was the same as in Experiment 1 (around 7 to 1), there was a general decrease in the retrieval of the critical sources. Collapsing across conditions, the mean retrieval rate of critical sources dropped from 42.71% (Experiment 1) to 15.91% (Experiment 2). One possible reason for this general drop-off could be that episodes corresponding to a single SCG (Experiment 1) constitute units involving short time slices of experience, whereas events encompassed by two connected SGCs (Experiment 2) may not. Hence, some of the events reported by participants of our second experiment might have not been originally encoded in episodic memory as a whole, being arranged as units during the retrieval process itself (episodic memories tend to represent relatively brief experiences; Anderson & Conway, 1993). To illustrate, a participant might have experienced an instance of making an effort related to her studies (e.g., studying for a quiz during regular vacations), which ultimately led to a high grade in the exam, but without having causally linked the outcome to the effort at the time of encoding. In cases like this, a possible retrieval strategy would consist of searching for instances of the SGC study-related effort, and, upon retrieving an instance, checking whether it was followed by an instance of the SGC study-related achievement. This kind of incremental search strategy, similar to the one implemented in O’Keefe and Costello’s (2008) model, may require a conscious effort, leading some participants to give up the search process and therefore fail to retrieve an analogous case. In favor of this interpretation, the proportion of participants failing to provide any kind of response during the second phase was much higher in Experiment 2 (45.5%) than in Experiment 1 (21.9%).

It could be the case that the SGCs of some of the cause-effect pairs are so strongly associated with one another that they could be considered to constitute a larger SGC (e.g., effort-related achievement). In cases like these, two events pertaining to such SGCs would be naturally interpreted as maintaining a causal relation, and encoded as such in LTM. If this were the case, retrieving the base episodes would be easier than when no such larger SGCs exist, and the kind of incremental search described above would not be necessary. Even though this represents an interesting avenue for inquiry that could attest to our explanation of the general drop-off in analogical retrieval that took place in our second experiment, the low retrieval rates in both conditions prevent us from assessing whether those sets of materials in which the causes and the effects described by participants seemed to be more strongly associated in fact behaved differently from the other sets.

As stated in the Introduction, an analysis of participants’ responses to free-production tasks is ill-suited for informing the tendencies of the memory system, since the proportions of superficially similar and superficially dissimilar items that are retrieved may to a great extent reflect the proportions of these two kinds of items that are potentially available for retrieval. Likewise, the hybrid paradigm employed in the present study should preclude deriving any inferences from reported LTM items that had not been previously detected via the availability questionnaire of the first phase. Hence, all conclusions originating from these data must be taken as utterly speculative. Across experiments and conditions, while 13% of participants retrieved a non-critical base analog that maintained surface similarities with the target, 2.7% of participants retrieved a non-critical base analog that did not share surface features with the target. With regard to non-analogical retrievals, while 7.6% of participants in Experiments 1 and 2 retrieved superficially similar items, 4.3% retrieved items that did not share surface features with the target (across experiments and conditions, 8.7% of responses were not classifiable as retrievals). These data can be taken to suggest that the retrieval rates of the critical distant analogs might have been negatively affected by the retrieval of analogous as well as non-analogous items maintaining surface similarity with the target.

Throughout the present study, we have contended that comparisons between exemplars of SGCs constitute a frequent case of daily-life analogizing. However, it could be argued that the observed prevalence of close analogs among participants’ responses could have originated in the fact that, for the particular task and materials employed in our experiments, participants might have been reluctant to regard comparisons to superficially distant exemplars of the target SGC as constituting valid analogies. To discard this possibility, we conducted an ad-hoc study with an independent sample of participants. We selected the 2 sets from Experiment 1 (sets 1 and 4) and the 2 sets from Experiment 2 (sets 3 and 4) that had elicited the lowest rates of distant retrievals. For each of these sets, we randomly selected a close and a distant response from those responses we had coded as belonging to the critical SGCs of Experiments 1 and 2, and paired each of them with its corresponding target situation. Thirty students from the same population as Experiments 1 and 2 received these eight pairs in random order, and interleaved with four filler pairs consisting of non-analogous episodes. All pairs were edited and presented as follows: “One person made the following comparison: When told that another person arrived late to take a plane, she compared the situation with an episode she had experienced, where she arrived late for a lung tomography. To what extent do you consider that these two situations are analogous? (1 = not analogous, 5 = very analogous)”. Comparisons between SCG exemplars received higher scores of analogical relatedness than between fillers (M = 3.73; SD = .51 vs. M = 1.84; SD = .34, t(29) = 17.4, p < .001), and so did comparisons between systems of SGC exemplars (M = 3.78; SD = .45 vs. M = 1.84; SD = .34, t(14) = 21.7, p < .00). More importantly, intradomain analogies did not differ from interdomain analogies either among comparisons between SGC exemplars (M = 3.78; SD = .61 vs. M = 3.68; SD = .62, t(29) = .82, p > .42) or among comparisons between systems of SGC exemplars (M = 3.85; SD = .58 vs. M = 3.71; SD =.61, t(29) = .94, p > .35). These results are consistent with studies dealing with analogy evaluation (Tavernini et al., 2017) and analogical re-representation (Oberholzer et al., 2018), in which participants granted high ratings of analogical relatedness to distant exemplars of SGCs. Hence, the predominance of superficially similar analogs in Experiments 1 and 2 cannot be attributed to participants’ reluctance to admit semantically distant exemplars of the target SGC as constituting valid sources for analogical comparisons. These results are also consistent with those of previous studies that employed alternative ways of eliciting judgments of analogical relatedness (e.g., Gentner et al., 1993; Gentner & Kurtz, 2006). Finally, the high ratings obtained by our analogies (around 3.7 in all conditions) demonstrates that the comparisons included in our materials are still considered analogous despite the simplification involved in shifting from real situations (e.g., two actual cases of corruption) to very succinct and decontextualized descriptions of hypothetical situations (e.g., to compare arriving late to take a plane to arriving late for a lung tomography).

General discussion

Authors like Dunbar (2001) and Hofstadter and Sander (2013) have considered that the low rates of distant retrieval obtained in traditional experimental studies have led to the wrong conclusion that distant retrieval is difficult at large. They argued that the experimental results could be attributed to the fact that the source analogs employed do not seem appropriate for eliciting the kind of abstract schemas that could facilitate their retrieval during the processing of a distant target. Under these adverse conditions, surface similarities become critical for analogical retrieval. According to these authors, the traditional conclusion that distant retrieval is generally unlikely may have resulted from overlooking the fact that lay-people can take advantage of overlearned schemas such as lexicalized SGCs (e.g., party, negotiation, or betrayal) to make sense of everyday experience. The application of the same SGC to two events would promote the highlighting and abstraction of the relational structure that renders these two events instances of the category. These categorizations could also lead to de-emphasizing the idiosyncratic surface content of exemplars, in turn facilitating the retrieval of stored items despite superficial mismatches with the target. Thus, the application of a lexicalized SGC would clearly represent a more favorable condition for remote transfer, as compared to the materials typically employed in the experimental tradition (Jamrozik & Gentner, 2020; Kurtz & Honke, 2020).

Contrary to these sensible expectations, in the present study we found a strong effect of surface similarities on the retrieval of stored exemplars of lexicalized SGCs. This effect was found for simple events that were cases of a single SGC (Experiment 1), as well as for comparatively more complex situations involving a causal connection between instances of two different SGCs (Experiment 2). These results contradict those of previous studies (e.g., Olguín et al., 2017a; Raynal et al., 2018), whose free-production paradigm suffers from a number of methodological shortcomings.

The obtained results are surprising, given the well-established and widespread finding that promoting an abstract encoding of base and target analogs facilitates distant retrieval (see Trench & Minervino, 2020, for a review). Traditional studies have repeatedly found that people do not spontaneously derive the type of schemas that support distant transfer. Based on this fact, a long tradition of interventions has been developed to increase distant retrieval by way of promoting an abstract encoding of base analogs. The most robust intervention consists in promoting schema induction by asking participants to compare two superficially dissimilar base analogs. To exemplify, in a pioneering study Gick and Holyoak (1983) found that while 45% of participants in the comparison condition spontaneously generated the convergence solution for the tumor problem, only 21% of participants in the single analog condition did so. Catrambone and Holyoak (1989) replicated these results for conditions in which the contextual separation between the learning and the transfer phases was stronger. Apart from comparison, other types of interventions directed to promote an abstract encoding of base analogs have also proven successful (e.g., abstracting a schema from a single example, generating analogous examples, and shifting for concrete to idealized representations; see Trench & Minervino, 2020, for a review).

Kurtz and Loewenstein (2007) reasoned that if retrieval depends on the degree of match between a stored item and a target situation, the beneficial effect of schema abstraction should also apply when elaborating on the target situation at retrieval time. Asking participants to compare two superficially dissimilar target problems proved effective in increasing the retrieval of a superficially dissimilar analog. The employment of the convergence strategy under this condition (54%) was higher than under a baseline condition receiving only the tumor problem (15%). Gentner et al. (2009) replicated this finding despite having enforced a stronger contextual separation between the learning and the transfer phases. As with base elaborations, other interventions different from comparison have also proven effective in increasing relational transfer (see Trench & Minervino, 2020, for a review).

The assumption underlying both research programs is that the best condition for transfer is one wherein a person has encoded both the base and the target in terms of the abstract schema they shared, well-established lexicalized categories being especially appropriate (Jamrozik & Gentner, 2020). However, Kurtz and Honke (2020) have manifested some frustration about what has been achieved by these interventions, since the obtained rates of distant retrieval (from one-quarter to one-half of participants) are still far from ceiling performance. In their own words: “Is it really so hard for people to notice that a new problem is fundamentally like the ones they just saw? Is there a fundamental limitation intrinsic to the cognitive system – or are researchers somehow not thinking about this right?” (p. 804). Without doubt, these authors would find the present findings somewhat astonishing: the levels of distant retrievals that we obtained in our experiments are much lower than those typically observed in studies of analogical reminding in which participants are struggling to encode one of the analogs in terms of a schema that they are learning for the first time in sessions of 20 min or so. How do we explain the difference between the results of these educationally oriented experiments and the present, more naturalistic study wherein participants are proficient users of the exact lexicalized schemas that are needed to uniformly categorize both the base and the target?

In order to ward off criticisms pointing to the lack of personal significance of typical laboratory materials (Blanchette & Dunbar, 2000; Hofstadter & Sander, 2013), we evaluated the retrieval of daily-life autobiographical episodes. Albeit familiar to participants, it is possible that autobiographical episodes might not be ideally suited for demonstrating the agreed-upon benefits of encoding the base and the target in terms of an overlearned schema. The kind of events that are normally stored in episodic memory might not be particularly amenable to de-emphasizing surface content in favor of an applicable relational category. As episodic memory is the memory of everyday events, it preserves information about the times, location, and associated emotions, as well as other contextual who-what-when-where-why knowledge that can be explicitly stated or conjured. This way, although cementing a new episodic memory requires its processing in terms of concepts coming from semantic memory (Tulving, 2002; e.g., a SGC is applied to a particular event), this not necessarily entails suppressing the details of the experience, since it is these details that constitute the core substance for the episode to be revived in a figurative travel back in time. To exemplify, think of a situation wherein a thief stole a watch by way of threatening you with a bladed weapon. Will you discard the details of the event due to having encoded it as an instance of robbery? Will this episode be just as likely to be activated after stealing a bracelet while brandishing a gun as after stealing a wallet from your bag? In fact, laboratory studies credited with having augmented distant retrieval by promoting an encoding of the base and/or the target in terms of schemas, have worked with materials associated with semantic memory. Future studies could delve into the differential effect of abstract encodings as a function of whether the stored items are of semantic or episodic nature.

Another possible reason behind our low rates of distant retrieval might concern the availability of near analogs that could outcompete the distant analog whose retrieval was being investigated. As revealed by the non-critical analogical events retrieved by participants, most responses came from the same domain as the target. The success of laboratory interventions in promoting the retrieval of distant sources may likely stem from the fact that such analogs did not compete with near analogs in the LTM of participants. If the military story were stored in LTM along with many near analogs of the tumor problem, its chances of being retrieved would drop significantly, even under optimal conditions of encoding. As computationally simulated by Gentner et al. (2009), the observed retrieval advantage of more generalized versions of the target were no longer observed in those cases where the distant source competed with a close analog in LTM. Future behavioral studies should aim at detecting whether the encoding of events in terms of SGCs aids distant retrieval in those cases where there is no competition of near analogs in LTM. This kind of retrieval will become necessary whenever a very novel instance of a SGC (e.g., a very strange case of addiction) does not have a superficially similar antecedent in our culture.

The difficulties involved in spotting a distant base analog among a wealth of intradomain analogs should not be regarded as a matter of concern, since retrieving literally similar sources generally represents a better choice than retrieving superficially dissimilar analogs. The fact that two situations maintain a wide array of perceivable resemblances increases the probability that less salient structural features will also be shared (the kind world hypothesis; Gentner et al., 1993). To illustrate, imagine that you are to persuade somebody else that ingesting too much passion-fruit will make her stop liking it. Even though analogies with (a) somebody who got sick of eating too many lychees, and (b) somebody who got tired of playing too many computer games, might probably obtain similar scores on traditional measures of analogical soundness, it seems clear that by relying on the same set of biological foundations, the precedent of having been fed-up with lychees confers greater probability to the conclusion. Our result that the retrieval advantage of near over distant analogs did not disappear when base and targets were cases of overlearned schemas may relate to the superiority of close analogs for activities as evolutionarily relevant as prediction. If we were to undergo the extraction of a tumor in our stomach, we would hardly congratulate our physician for delving into books of military strategy instead of reviewing similar medical interventions. We would always prefer that experts exhaust their intradomain cases before trying unnecessary mental leaps. This is not to say that distant analogs would not be required under particular pragmatics and pressures, such as building an argument or an explanation for an audience that is not familiar with the target topic. As suggested by recent results (e.g., Olguín, Trench & Minervino, 2017b; Trench et al., 2016), our surface bias in analogical retrieval could be somewhat alleviated by means of a voluntary and strategic kind of search.