The hippocampal-entorhinal system represents nested hierarchical relations between words during concept learning

A fundamental skill of an intelligent mind is that of being able to rapidly discover the structural organization underlying the relations across the objects or the events in the world. Humans, thanks to language, master this skill. For example, a child learning that dolphins and cats can also be referred to as mammals, not only will infer the presence of a hierarchical organization for which dolphins and cats are subordinate exemplars of the category mammals, but will also derive that dolphins are, at least at one conceptual level, more similar to cats than to sharks, despite their indisputable higher perceptual similarity to the latter. The hippocampal-entorhinal system, classically known for its involvement in relational and inferential memory, is a likely candidate to construct and hold these complex relational structures between concepts. To test this hypothesis, we trained healthy human adults to organize a novel audio-visual object space into categories labeled with novel words. Crucially, a hierarchical taxonomy existed between the object categories, and participants discovered it via inference during a simple associative object-to-word training. Using functional MRI after learning, and a combination of ROI-based multivariate analyses, we found that both the mid-anterior hippocampus and the entorhinal cortex represented the inferred hierarchical structure between words: subordinate-level words were represented more similarly to their related superordinate than to unrelated ones. This was paired, in the entorhinal cortex, by an additional signature of internalized structural representation of nested hierarchy: words referring to subordinate concepts belonging to the same superordinate category were represented more similarly compared with those not belonging to the same superordinate level: interestingly, this similarity was never directly taught to subjects nor it was made explicit during the task, but only indirectly derived through a logical inferential process and, crucially, contrasted the evidence coming from the definitional perceptual properties of the concepts. None of these results were observed before learning, when the same words were not yet semantically organized. A whole-brain searchlight revealed that the effect in the entorhinal cortex extends to a wider network of areas, encompassing the prefrontal, temporal, and parietal cortices, partially overlapping with the semantic network.