Poster Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions

Encoding of Grammatical Case in Human Hippocampus

Poster Session D, Saturday, September 13, 5:00 - 6:30 pm, Field House
This poster is part of the Sandbox Series.

James Belanger1,2, Melissa Franch1, Ana Chavez1, Raissa Mathura1, Steven Piantadosi3, Elizabeth Mickiewicz1, Kalman Katlowitz1, Suzanne Kemmer2, Benjamin Hayden1,2; 1Department of Neurosurgery, Baylor College of Medicine, 2Department of Linguistics, Rice University

Grammar allows us to compose complex meanings out of simpler lexical elements. Formally, operations like case declension alter semantic representations that modify meaning while preserving features of the underlying concept. To understand the neural basis of case declension, we analyzed responses of several hundred hippocampal neurons in humans listening to narrative English speech. We found that case declension corresponds to a specific and consistent translation of the population semantic vector. Critically, these effects persist even after controlling for confounds such as semantic features that predict case (e.g., animacy and agentivity), word order, syntactic depth, word duration, opening nodes, closing nodes and f0. Moreover, the brain uses similar rotations for noun and pronoun declension, indicating that these processes are at least somewhat consistent across semantic category. Finally, we found that the brain uses specific distinct semi-orthogonal subspaces for different grammatically marked forms; this coding scheme allows for unambiguous semantic shift while maintaining referential consistency. We confirm these results also reflect how these grammatical compositions work in embeddings within layers derived from LLMs. Together these results offer a specific neural coding-based solution to the problem of grammatical composition.

Topic Areas: Speech Perception, Syntax and Combinatorial Semantics

SNL Account Login


Forgot Password?
Create an Account