Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Persistent Homology Reveals Distinct Lexical-Semantic Geometry in Angular and Fusiform Gyri
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Zhonghua Tang1, Bailu Si1; 1School of Systems Science, Beijing Normal University, Beijing , China
How lexical-semantic knowledge is organized across cortical systems remains central to the neurobiology of language. The angular gyrus (AG) has been linked to semantic integration and cross-modal representation, whereas the fusiform gyrus (FG) supports visually grounded object and category information. We asked whether these regions differ in the higher-order organization of their conceptual representational spaces, beyond pairwise similarity. We reanalyzed two fMRI datasets from Fernandino et al. (2022), in which participants made familiarity judgments on visually presented English nouns (Study 1: 300 concepts; Study 2: 320 object/event concepts). Using the published subject-level, region-wise representational similarity matrices, we converted them to dissimilarity matrices and applied topological data analysis. Persistent-homology-derived Betti curves across increasing neighborhood density were summarized as Betti-curve integrals for homology dimensions H1–H3, indexing increasingly higher-order topological structure. Across both datasets, FG showed consistently stronger topological structure than AG. In Study 1 (N = 8), AG integrals (M ± SD) were 6.57 ± 1.21 (H1), 5.36 ± 3.47 (H2), and 3.67 ± 3.53 (H3); FG integrals were 20.48 ± 7.89, 59.50 ± 38.37, and 116.29 ± 104.88, with mean differences of 13.92 (95% CI [7.60, 20.23]), 54.14 ([22.85, 85.43]), and 112.62 ([25.74, 199.49]). Paired t tests showed FG > AG for H1, t(7) = 5.21, Bonferroni-corrected p = .0037, dz = 1.84; H2, t(7) = 4.09, p = .0139, dz = 1.45; and H3, t(7) = 3.07, p = .0546, dz = 1.08. All eight participants showed the same direction; exact Wilcoxon signed-rank and exact paired sign-flip permutation tests (all 2⁸ = 256 sign assignments enumerated) both yielded two-tailed p = .0078 for H1–H3. In Study 2 (N = 36), the dissociation replicated robustly. AG integrals were 8.11 ± 2.63, 8.83 ± 5.19, and 6.87 ± 5.40; FG integrals were 19.11 ± 3.66, 42.62 ± 15.95, and 62.50 ± 32.48, with mean differences of 11.01 (95% CI [9.41, 12.60]), 33.79 ([27.96, 39.62]), and 55.63 ([44.34, 66.92]). Paired t tests: H1, t(35) = 14.04, Bonferroni-corrected p = 1.79 × 10⁻¹⁵, dz = 2.34; H2, t(35) = 11.77, p = 3.05 × 10⁻¹³, dz = 1.96; H3, t(35) = 10.00, p = 2.52 × 10⁻¹¹, dz = 1.67. Effects survived FDR-BH correction, Wilcoxon signed-rank tests (W = 0 for all dimensions, Bonferroni-corrected p = 8.73 × 10⁻¹¹), and a one-million-iteration paired sign-flip permutation test (Bonferroni-corrected p = 3.00 × 10⁻⁶ for H1–H3). Together, these results indicate reliable AG–FG differences in the higher-order topology of conceptual representational spaces. The lower-amplitude AG profiles may reflect a smoother or more compressed representational geometry, consistent with its proposed role in integrative semantic processing. In contrast, the stronger FG profiles suggest a more heterogeneous geometry, potentially related to visual-category and object-property distinctions. These findings show that topology-sensitive analysis of subject-level representational similarity structure can reveal regional differences in conceptual organization that are not apparent from pairwise similarity measures alone.
Topic Areas: Methods, Meaning: Lexical Semantics