WavePhaseNet: A DFT-Based Method for Constructing Semantic Conceptual Hierarchy Structures (SCHS)

arXiv cs.CL · 6 hours ago

Researchers propose WavePhaseNet, a method using Discrete Fourier Transform to construct semantic hierarchy structures in language models by decomposing embeddings into frequency bands and reducing dimensionality from 24,576 to approximately 3,000 dimensions. The approach achieves this 92% reduction in embedding dimensions while preserving semantic meaning by leveraging spectral analysis and Zipf's law. The method applies cohomological regularization to control consistency and suppress hallucinations through graph-based loss functions.