Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Semantic representations in the narrative temporal processing hierarchy : insights from simulation of human narrative processing.
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Peter Ford Dominey1; 1INSERM U1093 CAPS - Cognition, Action and Plasticity Sensorimotor
Research in narrative neuroscience has revealed multiple manifestations of a temporal processing hierarchy in the brain during narrative processing. Input driven areas display fast processing time constants, with a progression in extension of time constants in successively higher cognitive areas. The existence of this hierarchy poses multiple questions including what is its origin, and how is it related to semantic processing. In response to the first question, we recently investigated how the interaction of input structure and network architecture can produce temporal processing hierarchies. We simulate human narrative processing using recurrent neural networks with two notable features: First: the input to the networks is a trajectory of word embeddings derived from narratives used in fMRI studies. The resulting network activity is a proxy for human fMRI activation. Second: the connectivity structure of the networks is constrained by naturalistic physical properties. These properties include an exponential distance rule (EDR) governing connectivity, such that the probability of neurons being connected decreases exponentially with distance. This produces a flow of information through the network, naturally creating a temporal processing hierarchy, simulating aspects of the human temporal processing hierarchy (Chang et al. PNAS 2022, 2025). We also used human connectome matrices to define the network connectivity (Triebkorn et al. Comm Biol 2025). This allows for simulation of the temporal processing hierarchy as observed by Chien and Honey (Neuron 2020). The current research investigates how the temporal processing hierarchy is related to semantic processing. Methods: We evaluated narrative processing across 40 distinct reservoir seeds (1,000 neurons each) driven by 8 natural language narratives, using a spatially constrained EDR model and a human-derived structural Connectome matrix model. Narrative inputs were provided to the first 300/1000 neurons. To measure the dimensionality across the networks, we calculated the participation ratio of neural activity traces across ten sequential processing layers. To evaluate semantic representation across these manifolds, we conducted cross-reservoir leave-one-out cross-validation (LOOCV) decoding. Results: Both the EDR and Connectome models generated gradient of manifold rank reduction across the processing hierarchy from the early input-driven layers (PR ~= 22) down to the deepest layers (PR ~= 7). This argues that the temporal processing hierarchy is associated with a network rank hierarchy. In LOOCV testing across the 40 seeds, the Connectome activity manifolds yielded robust cross-decoding accuracy for the 8 narratives, outperforming the EDR pipeline especially in the more distant areas. This difference is due to the presence of the white matter pathways that allow the deeper areas to have access both to the successive integration via local connections, and to higher dimensional activity. Conclusion: These findings argue that the narrative temporal processing hierarchy is the reflection of a neural architecture that provides semantic integration over multiple timescales. They suggest that the human connectome acts as a multi-scale spatial filter. It progressively funnels high-dimensional inputs into low-dimensional, long-timescale functional manifolds, that provide a geometric substrate optimized for robust narrative comprehension.
Topic Areas: Computational Approaches, Meaning: Discourse and Pragmatics