Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A biologically inspired model for the incremental representation of syntactic constituent structure in the human brain
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Nosrat Mohammadi1, Nina Kazanina1, Itsaso Olasagasti1; 1Department of Basic Neurosciences, Faculty of Medicine, University of Geneva, Geneva, Switzerland
Extracting the meaning of a sentence requires the inference and representation of its underlying syntactic structure, yet how neural populations in the human brain encode incremental structure building during language processing remains elusive. Here, we propose a psycholinguistically grounded and biologically inspired neural model of syntactic structure representation during comprehension. Several models for structured symbolic representations in vector spaces exist (Plate, 1995; Smolensky, 1990; Shastri & Ajjanagadde, 1993; Papadimitriou et al., 2020; Calmus et al., 2020); however, experimental support remains limited. Here, we take an alternative approach and consider that language processing in the brain might share in part representational strategies found in the prefrontal cortex (PFC) in primates and rodents performing structured tasks (e.g., Xie et al. 2022, El-Gaby et al. 2024). These studies have revealed that PFC contains orthogonal neural subspaces for each abstract task state (the rank in a sequence of saccade targets, Xie et al. 2022; progress within a four-location loop through a maze, El-Gaby et al. 2024), with each subspace encoding the corresponding saccade target or location on the maze. That is, for each abstract task state, there is a dedicated neural subspace capable of encoding an arbitrary item. Inspired by this work, we propose that the human brain organizes phrase structure in a similar way, by using dedicated and hierarchically organized subspaces for each syntactic category in a way that incorporates knowledge about the syntactic rules of a given language. We apply the framework to encode constituent structure, focusing on noun phrases (NPs) and prepositional phrases (PPs) in English as proof of principle and formalising them via simple phrase-structure rules: NP = (Det)+(Adj)^n+N; PP = Prep+NP; NP = NP+PP; parentheses denote optional categories and "^n" denotes an arbitrary number of an optional category. NPs and PPs already exhibit two key linguistic properties: "optionality", with the only obligatory element within the NP being the head= Noun; and "recursion", with an NP nestable within a larger NP. This work is currently in progress. In our planned implementation, structural connectivity between task-relevant subspaces incorporates linguistic knowledge and expectations (e.g., encountering an Adjective rules out the possibility of the next category being a Determiner, but creates expectations about the next item being a Noun or another Adjective). Ring attractor dynamics, in which stable states are arranged along a closed-loop continuous manifold, offer a way to implement recursion, while state progression along orthogonal dimensions allows for optionality. While our long-term goal is to derive testable predictions to falsify the model using existing intracranial neural data of humans listening to natural speech, our immediate goal is to present a working implementation capable of parsing NPs and PPs up to one level of recursion. This offers a proof of principle of a biologically grounded and mechanistically explicit account of some aspects of syntactic processing that bridges computational neuroscience and psycholinguistics.
Topic Areas: Syntax and Combinatorial Semantics, Methods