Structure Preservation in Semantic Compression of Language: A Preliminary Exploration Based on Brain-like Representation Models

Objective: In natural language processing and cognitive science, how high-level abstract semantics evolve from concrete sensorimotor experiences remains computationally underexplored. This study aims to utilize large language-brain-like prediction models to quantitatively investigate whether the representational structure of underlying physical actions is preserved during the “semantic compression” and dimensional elevation of language. Methods: This study constructed a “Semantic Topology Corpus” with strictly controlled syntactic structure and character length (comprising 10 motifs, 30 sentences total, covering physical action bases and their positive/negative abstract derivatives). Using the TRIBE v2 brain-like representation model, we predicted whole-brain activation 3D maps and quantified spatial topological features across different semantic levels through Representational Similarity Analysis (RSA) and significant activation voxel statistics (Z > 2.0 ). Results: (1) Physical actions and abstract derivative concepts demonstrated high whole-brain representational similarity (r ≈ 0.79 ), suggesting a cross-level “structure preservation effect”; (2) Abstract concepts with the same Level of Detail (LOD) but opposing emotional valences showed even higher similarity (r = 0.847 ), indicating a trend of representational clustering in high-dimensional semantics; (3) Voxel statistics revealed that during spatial topological transfer of abstract semantics, the total volume of significantly activated voxels did not exhibit substantial expansion (maintained at approximately 9,700 voxels), suggesting efficient reuse of representational subspace. Conclusion: The semantic compression process of language may not be a de novo reconstruction, but rather reuse and preserve the structure of underlying embodied actions in model-predicted topological space with high representational efficiency. This study provides quantitative evidence from computational models that is consistent with “Embodied Cognition” and Gestalt holistic processing theory, and offers insights for the underlying cognitive architecture of future Embodied AI and World Models.