Poster Presentation

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

A Geometric Foundation for Semantic Relations in the Brain

Hanlin Zhu¹, Melissa Franch¹, Elizabeth Mickiewicz², James Belanger¹, Rhiannon L Cowan³, Kalman Katlowitz¹, Ana Chavez¹, Assia Chericoni¹, Danika Paulo¹, Xinyuan Yan¹, Shervin Rahimpour³, Ben Shofty³, Eleonora Bartoli¹, Jay Hennig¹, Nicole Provenza¹, Elliot H Smith³, Steven Piantadosi⁴, Sameer A. Sheth¹, Benjamin Y Hayden¹; ¹Baylor College of Medicine, ²Pennsylvania State University, ³University of Utah, ⁴University of California, Berkeley

Presenter: Hanlin Zhu

In vectorial word embeddings, semantic features often appear as reusable directions: the same gender direction can distinguish “boy”/”girl”, “king”/”queen”, and “uncle”/”aunt”. Here we show that semantically driven neural responses in the human brain follow a similar geometric principle during natural language comprehension. We recorded populations of single neurons from hippocampus, anterior cingulate cortex, and orbitofrontal cortex while participants listened to podcasts. Single neurons often discriminated the two poles of a semantic relation, combining these relation-tuned neurons yielded population embeddings whose shared displacements generalized across word families. Across diverse semantic and grammatical relations, words linked by the same relation occupied aligned directions in neural population space, yielding parallelogram structure. Pronouns showed a stronger higher-order organization, with person, number, and case combining into an approximately prismatic geometry consistent with partial factorization. Departures from ideal parallelism resembled those in a modern language model, suggesting a shared fine-grained semantic geometry. Different analogy types recruited largely distinct neuronal subpopulations, comparisons across brain regions revealed complementary specialization rather than a single uniform code. Together, these findings suggest that word meaning in the brain is organized by reusable geometric transformations, paralleling principles seen in artificial language systems.

Topic Area: Methods, Tools, Theory & Neural Coding