Relational labeling unlocks inert knowledge
Introduction
Vocabulary learning is often portrayed as a pointless exercise assigned to keep students busy. Yet learning certain kinds of terms may be extremely valuable for achieving domain mastery. The rationale for this is that higher-order cognition depends on relational processing—on the ability to represent and reason about relations such as causation and prevention in science, commutativity and distributivity in mathematics, and promise and lie in social interactions. We suggest that learning and using terms that denote these relational patterns is instrumental in acquiring cognitive flexibility and insight in and across domains.
In this paper we focus on one specific way in which language can support cognition— relational retrieval from long-term memory. Decades of research in the laboratory have shown poor retrieval of relational matches from memory. Given a current situation, people are often reminded of prior situations that share specific content features—objects, characters, locations, and associated entities—and fail to retrieve those that share relational structure (Brooks, 1987; Brooks, Norman, & Allen, 1991; Forbus, Gentner, & Law, 1995; Gentner, Rattermann, & Forbus, 1993; Holyoak & Koh, 1987; Ross, 1987, Ross, 1989; Trench & Minervino, 2015). This is true even when the relational match is demonstrably stored in memory (Gentner et al., 1993; Gick and Holyoak, 1980, Gick and Holyoak, 1983; Holyoak & Koh, 1987; Trench & Minervino, 2015); and it holds even when the same people later rate the relational matches (that did not come to mind) as both more similar and more inferentially sound than the surface matches that they readily retrieved (Gentner et al., 1993).
One way this pattern of rare relational remindings has been interpreted is in terms of differential encoding variability between relational information and surface contextual information about the entities that occupy roles within a relational structure (Forbus et al., 1995; Gentner, Loewenstein, Thompson, & Forbus, 2009). The idea is that relational information is often encoded in a context-specific way, and is therefore variably encoded across situations. In contrast, the entities within this context—such as objects, animals, and characters—tend to be uniformly encoded (Asmuth & Gentner, 2017; Bassok, Wu, & Olseth, 1995; Forbus et al., 1995; Gentner & France, 1988; Gentner et al., 2009;). An example of this phenomenon comes from work on verb mutability (Gentner, 1981; Gentner & France, 1988; Reyna, 1980). The term mutability refers to a word's propensity to assume different meanings across varying contexts. Evidence that verbs (which typically name relations) are more mutable than concrete nouns (which typically name object or animal categories) comes from studies by Gentner and France (1988) in which people were asked to paraphrase semantically strained sentences such as “The car worshipped.” The dominant response was to use a synonym for the noun (roughly preserving the noun's usual referent) and to alter the meaning of the verb to fit that referent: e.g., “The vehicle only responded to him,” or “Someone's vehicle was given a rest on a Sunday.”
Verb mutability has implications for the encoding and recognition of verbs versus concrete nouns (Earles & Kersten, 2017; Kersten & Earles, 2004; King & Gentner, 2019). For example, Earles and Kersten (2017) gave people simple sentences to remember, and later tested them with different combinations of words. In the test, they were asked to recognize a specific target word—either a verb or a noun—within each test sentence. Overall, nouns were better recognized than verbs. More to the point, verbs, but not nouns, showed a significant deficit from change of context. People were worse at recognizing verbs when they were paired with new nouns than when they appeared with the original nouns—for example, if the original was “drop the spoon,” people were better able to recognize “drop” in that same phrase than in a new phrase “drop the photograph”. In contrast, for nouns, recognition was equally good whether they were paired with different verbs or the same verb—for example, people were equally able to recognize “spoon” in the new phrase “bend the spoon” as in the previously-seen phrase “drop the spoon” (Earles & Kersten, 2017). These findings support Gentner's (1981) verb mutability hypothesis. Because the verbs were encoded differently in the new context than in the original context, the likelihood of experiencing a match between the test sentence and the input sentence was relatively low. In contrast, the nouns were likely to be encoded in the same way in both contexts, making it likely that people would experience a match.
Forbus et al. (1995) proposed that more generally, relations tend to be encoded in a context-specific way. And because spontaneous reminding from long-term memory depends on a match between the current situation and a stored representation, elements that are uniformly encoded (such as objects and characters), have an advantage in reminding over relational patterns, which are variably encoded.
Despite much evidence demonstrating poor relational retrieval, people do sometimes retrieve purely relational matches, with little or no surface support (Gentner et al., 1993). In the world outside the lab, creative solutions in science, design, and technology often come about because of spontaneous analogies, which rely on relational mappings within and between domains (e.g., Dunbar, 1995; Hargadon & Sutton, 1997; Majchrzak, Cooper, & Neece, 2004). For instance, designers at Nike developed a shock-absorbing shoe by drawing on a solution from Formula One race car suspension (Kalogerakis, Lüthje, & Herstatt, 2010). There is evidence that relational remindings become more likely as people gain in domain knowledge (Goldwater & Schalk, 2016; Goldwater, Sibley, Gentner, LaDue, & Libarkin, under review; Koedinger & Roll, 2012; Novick, 1988). This cannot be due to simple accrual of examples, because this would also permit many more potential surface matches. A possible explanation, based on the above discussion, is that as people gain in domain knowledge, their representations become more sophisticated; they come to encode domain phenomena in terms of a set of higher-order relational schemas that apply widely in the domain (such as positive feedback loop or conservation of energy) (Forbus et al., 1995; Gentner et al., 2009; Goldwater & Gentner, 2015). The habitual use of key relational patterns during encoding promotes uniform relational representation, making it likely that a current example will overlap relationally with an example or abstraction stored in memory. Uniform relational representation could thus contribute to domain experts' superior relational retrieval.
How might experts' relational uniformity come about? There could be many contributing factors. People's self-explanations may contribute to deeper, more consistent representations of a domain (Chi, Feltovich, & Glaser, 1981; Legare & Lombrozo, 2014; Lombrozo, 2016). Analogical comparisons (whether guided or discovered) can highlight a common system of relations, inviting a more uniform representation across the two analogs (Catrambone & Holyoak, 1989; Day & Asmuth, 2017; Doumas & Hummel, 2013; Gentner et al., 2009; Gentner & Christie, 2010; Gick & Holyoak, 1983; Loewenstein, Thompson, & Gentner, 1999). We propose another factor that can contribute to uniform relational encoding—learning the relational language that characterizes a domain. We focus particularly on terms that name relational categories—for example, carnivore or positive feedback loop. Relational categories are categories whose members do not in general share common intrinsic features; rather, category membership is based on a common relational pattern (Gentner & Kurtz, 2005; Goldwater & Schalk, 2016; Kurtz, Boukrina, & Gentner, 2013; Markman & Stilwell, 2001). For example, members of the relational schema category positive feedback loop include the increasing audio feedback resulting from a speaker being placed too close to a microphone; increasing global temperatures resulting from the melting of polar ice; and increasing contractions during childbirth resulting from the release of oxytocin. We suggest that applying a relational schema category label such as positive feedback loop to an example invites structuring the example in terms of the overarching schema. Habitual use of the relational vocabulary of a domain could thus promote uniform relational representation across examples, and thereby increase relational retrieval.
This account predicts that applying a relational schema label to an example during initial encoding should increase the likelihood of later relational retrieval of that example. Less obviously, it also predicts increased relational reminding of prior relationally similar examples if a relational label is applied at test time. This second prediction follows from evidence that deriving a schema (by comparing examples of a relational structure) at retrieval time improves retrieval of prior relationally similar examples (Gentner et al., 2009; Kurtz & Loewenstein, 2007)—a phenomenon dubbed ‘late abstraction’. Likewise, if relational schema labels invite a corresponding relational construal, their use should promote retrieving prior examples that were encoded with the same construal. Of course, providing labels at both encoding and test should be especially effective; but this outcome will not be definitive, since it could result from the labels acting as a purely lexical cue that the two passages match (because they have the same label), as well as (or instead of) a relational match.
Across two studies, we used a cued-recall paradigm to test these predictions. Participants studied one set of passages in an encoding phase. After a delay, in the test phase they received a new set of passages, and were told to write any encoding passages of which they were reminded. Each test passage had the same relational structure as one of the original passages (e.g., positive feedback loop). In addition, to capture the challenge of real-life memory retrieval, for each test passage there was another original passage from the same domain as the test passage (e.g., medicine). Because the same-domain passage shared common contextual features—objects, characters, locations, and associated entities—with the test passage, we expected it to be a potent retrieval match. Thus, for each test passage there were two potential retrieval candidates: the relational match and the domain match (Fig. 1). In the absence of labels, we expected relational retrieval to be low and domain retrieval to dominate. The predictions are that (1) in the baseline condition, with no relational labels, domain retrievals will dominate; (2) providing relational labels at encoding will improve relational retrieval; and (3) this effect will also hold if relational labels are provided at test. A further prediction, tested in Experiment 2, is (4) that relational labels will lead to greater change in retrieval patterns relative to baseline than domain labels. This follows from our earlier claim that the specific content features that would be highlighted by domain labels are already encoded similarly across situations by default. We tested predictions (1), (2) and (3) in Experiments 1 and 2, and prediction (4) in Experiment 2.
In addition, in Experiment 1, we varied the placement of the label within a given passage. If relational labels serve to organize the passage according to a relational pattern, then they should be more powerful if given at the start of the passage (e.g., Ausubel, 1960; Bransford & Johnson, 1972).
Section snippets
Participants
Participants (N = 251, 154 female, mean age = 19.50) were native English speakers from the Northwestern University community who were paid or received course credit for participation. An additional 15 participants were tested but excluded from analyses for failing to follow the test instructions (3), for failing to respond to at least half of the test items (10), or due to experimental error (2). The sample size of 36 participants/condition was set based on expected effect sizes from prior
Experiment 1 Discussion
As predicted, relational labels markedly improved the likelihood of relational retrieval. Not surprisingly, having labels at both encoding and test was highly effective: people in this condition retrieved roughly four times as many relational matches as in baseline. More importantly, receiving relational labels at encoding was also highly effective. Compared to baseline, people were twice as likely to retrieve a relational match when given a relational label at encoding. Interestingly,
Participants
Participants (N = 114, 80 female, mean age = 20.82) were native English speakers recruited from the Northwestern University community. They were paid or received course credit for their participation. An additional five participants were tested but excluded from analyses for failing to respond to at least half of the test items. The sample size of 17 participants/condition was based on effect sizes observed in Experiment 1. This sample size had power of 80% to detect an effect the size of that
General discussion
We have advanced two claims. First, a major reason that relational retrieval is poor relative to retrieval based on surface similarity is differential encoding variability—information about objects and other entities is typically encoded uniformly across situations, whereas relational information is encoded in a context-specific way (Asmuth & Gentner, 2017; Forbus et al., 1995; Gentner et al., 2009; Kersten & Earles, 2004). Second, the use of a relational schema label invites a corresponding
Conclusion
Our findings show that relational labels can have a powerful effect on people's ability to encode and retrieve examples of relational patterns. We suggest that the use of relational labels highlights relational patterns that might otherwise be missed, or bound to the specific features of examples. We further suggest that habitual use of domain-relevant relational terms is a major contributor to experts' superior relational encoding and retrieval.
Author contributions
A. Jamrozik: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data Curation, Writing – Original Draft, Writing - Review & Editing.
D. Gentner: Conceptualization, Methodology, Resources, Writing - Review & Editing, Supervision.
Acknowledgments
This research was funded by ONR Grant N00014-08-1-0040. We thank Laura Willig for her help in coding the data, Becky Bui, Lizabeth Huey, and Benjamin Dionysus for their administrative support, and the Cognition and Language Lab for their helpful suggestions.
References (66)
- et al.
Adding apples and oranges: Alignment of semantic and formal knowledge
Cognitive Psychology
(1998) - et al.
Contextual prerequisites for understanding: Some investigations of comprehension and recall
Journal of Verbal Learning and Verbal Behavior
(1972) - et al.
Categorization and representation of physics problems by experts and novices
Cognitive Science
(1981) - et al.
The effects of manifest relational similarity on analog retrieval
Journal of Memory and Language
(1994) - et al.
MAC/FAC: A model of similarity-based retrieval
Cognitive Science
(1995) - et al.
The verb mutability effect: Studies of the combinatorial semantics of nouns and verbs
- et al.
Comparison in the development of categories
Cognitive Development
(1999) - et al.
The roles of similarity in transfer: Separating retrievability from inferential soundness
Cognitive Psychology
(1993) - et al.
Analogical problem solving
Cognitive Psychology
(1980) - et al.
Schema induction and analogical transfer
Cognitive Psychology
(1983)
On the acquisition of abstract knowledge: Structural alignment and explication in learning causal system categories
Cognition
Selective effects of explanation on learning during early childhood
Journal of Experimental Child Psychology
Relational language and the development of relational mapping
Cognitive Psychology
Explanatory preferences shape learning and inference
Trends in Cognitive Sciences
Words as invitations to form categories: Evidence from 12- to 13-month-old infants
Cognitive Psychology
The role of language in acquiring object kind concepts in infancy
Cognition
Relational categories are more mutable than entity categories
The Quarterly Journal of Experimental Psychology
The use of advance organizers in the learning and retention of meaningful verbal material
Journal of Educational Psychology
Using content to interpret structure: Effects on analogical transfer
Current Directions in Psychological Science
Judging a book by its cover: Interpretative effects of content on problem-solving transfer
Memory & Cognition
Decentralized control of categorization: The role of prior processing episodes
Role of specific similarity in a medical diagnostic task
Journal of Experimental Psychology: General
Overcoming contextual limitations on problem-solving transfer
Journal of Experimental Psychology Learning Memory and Cognition
Language helps children succeed on a classic analogy task
Cognitive Science
Re-representation in comparison and similarity
Comparison and mapping facilitate relation discovery and predication
PLoS One
How scientists really reason: Scientific reasoning in real-world laboratories
Why are verbs so hard to remember? Effects of semantic context on memory for verbs and nouns
Cognitive Science
Creative imitation: Exploring the case of cross-industry innovation
R&D Management
Some interesting differences between verbs and nouns
Cognition and Brain Theory
Why we're so smart
Bootstrapping the mind: Analogical processes and symbol systems
Cognitive Science
Language as cognitive toolkit: How language supports relational thought
The American Psychologist
Cited by (30)
Analogical inferences mediated by relational categories
2023, Cognitive PsychologyChildren's gesture use provides insight into proportional reasoning strategies
2022, Journal of Experimental Child PsychologyCitation Excerpt :Using visual displays of proportion, they introduced equivalent proportions (e.g., 3/4, 6/8) using either a single categorical label “blick” or traditional fraction labels of “three fourths” and “six eighths.” Based on prior work investigating the use of verbal labels for highlighting relations and category membership (Fulkerson & Waxman, 2007; Fyfe, McNeil, & Rittle-Johnson, 2015; Jamrozik & Gentner, 2020), Hurst and Cordes (2019) hypothesized that the single categorical label would help children attend to proportion and inhibit numerical interference relative to traditional labels. This is exactly what they found in two separate studies with 6-year-old children; seeing equivalent fractions labeled with the same categorical label helped children to inhibit whole number biases and attend to proportional information compared with hearing a traditional fraction label.
Do graded representations support abstract thought?
2021, Current Opinion in Behavioral SciencesCitation Excerpt :Our argument is that presence or absence of symbols is not itself a good account of comparative or developmental differences in same-different reasoning. Nevertheless, the relative accessibility of verbal symbols to adults could explain both why adults are so quick to recognize sameness and why they sometimes fail to do so [32] (Box 3). In sum, we believe the evidence should lead us to reject discontinuous symbolic accounts of relational reasoning.
Learning same and different relations: cross-species comparisons
2021, Current Opinion in Behavioral SciencesCitation Excerpt :However, once acquired, language supports human relational learning and reasoning in several ways [1,2,15,16]. First, having a symbolic label for a concept supports the ability to notice and use that concept, in children [11,12,23,24,15–19,20•,21,22] and adults [25,26•,27,28•]. Second, using a common label for two items prompts children to compare them, allowing them to notice and retain their commonalities [10,11,24,29•].
Natural language processing in-and-for design research
2022, Design ScienceTransfer of learning in histology: Insights from a longitudinal study
2024, Anatomical Sciences Education