Distributed Code for Semantic Relations Predicts Neural Similarity during Analogical Reasoning

Relational and lexical similarity in analogical reasoning and recognition memory: Behavioral evidence and computational evaluation

We examined the role of different types of similarity in both analogical reasoning and recognition memory. On recognition tasks, people more often falsely report having seen a recombined word pair (e.g., flower: garden) if it instantiates the same semantic relation (e.g., is a part of) as a studied word pair (e.g., house: town). This phenomenon, termed relational luring, has been interpreted as evidence that explicit relation representations-known to play a central role in analogical reasoning-also impact episodic memory. We replicate and extend previous studies, showing that relation-based false alarms in recognition memory occur after participants encode word pairs either by making relatedness judgments about individual words presented sequentially, or by evaluating analogies between pairs of word pairs. To test alternative explanations of relational luring, we implemented an established model of recognition memory, the Generalized Context Model (GCM). Within this basic framework, we compared representations of word pairs based on similarities derived either from explicit relations or from lexical semantics (i.e., individual word meanings). In two experiments on recognition memory, best-fitting values of GCM parameters enabled both similarity models (even the model based solely on lexical semantics) to predict relational luring with comparable accuracy. However, the model based on explicit relations proved more robust to parameter variations than that based on lexical similarity. We found this same pattern of modeling results when applying GCM to an independent set of data reported by Popov, Hristova, and Anders (2017). In accord with previous work, we also found that explicit relation representations are necessary for modeling analogical reasoning. Our findings support the possibility that explicit relations, which are central to analogical reasoning, also play an important role in episodic memory.

From Semantic Vectors to Analogical Mapping

Human reasoning goes beyond knowledge about individual entities, extending to inferences based on relations between entities. Here we focus on the use of relations in verbal analogical mapping, sketching a general approach based on assessing similarity between patterns of semantic relations between words. This approach combines research in artificial intelligence with work in psychology and cognitive science, with the aim of minimizing hand coding of text inputs for reasoning tasks. The computational framework takes as inputs vector representations of individual word meanings, coupled with semantic representations of the relations between words, and uses these inputs to form semantic-relation networks for individual analogues. Analogical mapping is operationalized as graph matching under cognitive and computational constraints. The approach highlights the central role of semantics in analogical mapping.

The Neural Correlates of Analogy Component Processes

Analogical reasoning is a core facet of higher cognition in humans. Creating analogies as we navigate the environment helps us learn. Analogies involve reframing novel encounters using knowledge of familiar, relationally similar contexts stored in memory. When an analogy links a novel encounter with a familiar context, it can aid in problem solving. Reasoning by analogy is a complex process that is mediated by multiple brain regions and mechanisms. Several advanced computational architectures have been developed to simulate how these brain processes give rise to analogical reasoning, like the "learning with inferences and schema abstraction" architecture and the Companion architecture. To obtain this power to simulate human reasoning, theses architectures assume that various computational "subprocesses" comprise analogical reasoning, such as analogical access, mapping, inference, and schema induction, consistent with the structure-mapping framework proposed decades ago. However, little is known about how these subprocesses relate to actual brain processes. While some work in neuroscience has linked analogical reasoning to regions of brain prefrontal cortex, more research is needed to investigate the wide array of specific neural hypotheses generated by the computational architectures. In the current article, we review the association between historically important computational architectures of analogy and empirical studies in neuroscience. In particular, we focus on evidence for a frontoparietal brain network underlying analogical reasoning and the degree to which brain mechanisms mirror the computational subprocesses. We also offer a general vantage on the current- and future-states of neuroscience research in this domain and provide some recommendations for future neuroimaging studies.

Analogy-Related Information Can Be Accessed by Simple Addition and Subtraction of fMRI Activation Patterns, Without Participants Performing any Analogy Task

Analogical reasoning, for example, inferring that teacher is to chalk as mechanic is to wrench, plays a fundamental role in human cognition. However, whether brain activity patterns of individual words are encoded in a way that could facilitate analogical reasoning is unclear. Recent advances in computational linguistics have shown that information about analogical problems can be accessed by simple addition and subtraction of word embeddings (e.g., wrench = mechanic + chalk - teacher). Critically, this property emerges in artificial neural networks that were not trained to produce analogies but instead were trained to produce general-purpose semantic representations. Here, we test whether such emergent property can be observed in representations in human brains, as well as in artificial neural networks. fMRI activation patterns were recorded while participants viewed isolated words but did not perform analogical reasoning tasks. Analogy relations were constructed from word pairs that were categorically or thematically related, and we tested whether the predicted fMRI pattern calculated with simple arithmetic was more correlated with the pattern of the target word than other words. We observed that the predicted fMRI patterns contain information about not only the identity of the target word but also its category and theme (e.g., teaching-related). In summary, this study demonstrated that information about analogy questions can be reliably accessed with the addition and subtraction of fMRI patterns, and that, similar to word embeddings, this property holds for task-general patterns elicited when participants were not explicitly told to perform analogical reasoning.

Individual Differences in the Neural Localization of Relational Networks of Semantic Concepts

Semantic concepts relate to each other to varying degrees to form a network of zero-order relations, and these zero-order relations serve as input into networks of general relation types as well as higher order relations. Previous work has studied the neural mapping of semantic concepts across domains, although much work remains to be done to understand how the localization and structure of those architectures differ depending on various individual differences in attentional bias toward different content presentation formats. Using an item-wise model of semantic distance of zero-order relations (Word2vec) between stimuli (presented both in word and picture forms), we used representational similarity analysis to identify individual differences in the neural localization of semantic concepts and how those localization differences can be predicted by individual variance in the degree to which individuals attend to word information instead of pictures. Importantly, there were no reliable representations of this zero-order semantic relational network when looking at the full group, and it was only through considering individual differences that a stable localization difference became evident. These results indicate that individual differences in the degree to which a person habitually attends to word information instead of picture information substantially affects the neural localization of zero-order semantic representations.

Transient Neural Activation of Abstract Relations on an Incidental Analogy Task

Although a large proportion of the lexicon consists of abstract concepts, little is known about how they are represented by the brain. Here, we investigated how the mind represents relations shared between sets of mental representations that are superficially unrelated, such as car-engine and dog-tongue, but that nonetheless share a more general, abstract relation, such as whole-part. Participants saw a pair of words on each trial and were asked to indicate whether they could think of a relation between them. Importantly, they were not explicitly asked whether different word pairs shared the same relation, as in analogical reasoning tasks. We observed representational similarity for abstract relations in regions in the "conceptual hub" network, even when controlling for semantic relatedness between word pairs. By contrast, we did not observe representational similarity in regions previously implicated in explicit analogical reasoning. A given relation was sometimes repeated across sequential word pairs, allowing us to test for behavioral and neural priming of abstract relations. Indeed, we observed faster RTs and greater representational similarity for primed than unprimed trials, suggesting that mental representations of abstract relations are transiently activated on this incidental analogy task. Finally, we found a significant correlation between behavioral and neural priming across participants. To our knowledge, this is the first study to investigate relational priming using functional neuroimaging and to show that neural representations are strengthened by relational priming. This research shows how abstract concepts can be brought to mind momentarily, even when not required for task performance.