Culturally-Guided Style Transfer for Southeast Asian Text using Constrained Diffusion Models
D. X. Cai, L. Chang · Advanced Electromagnetics · 2026
Preserving culturally grounded semantics during multilingual literary style transfer remains challenging because existing text generation models often fail to maintain narrative structure, symbolic imagery, and rhetorical consistency across heterogeneous linguistic contexts. This study proposes a culturally guided style transfer framework based on constrained diffusion models that jointly models semantic hierarchies and stylistic representations in a unified latent space. By introducing style-modifying latent variables into the diffusion process and coupling them with hierarchical semantic embeddings, the proposed approach enables synchronized regulation of cultural semantics and literary style throughout forward diffusion and reverse denoising. Unlike conventional conditional generation methods, stylistic evolution emerges through semantic-aware latent perturbations, improving the preservation of cultural identity while maintaining semantic coherence. Experiments on five Southeast Asian literary corpora demonstrate stable cultural mapping and consistent style reconstruction, achieving high levels of cultural image consistency together with reliable narrative and rhetorical preservation across multilingual settings. Beyond literary generation, the proposed semantic-driven diffusion framework offers a transferable paradigm for intelligent information representation and cross-modal content encoding, providing potential insights for semantic communication, electromagnetic information transmission, and AI-assisted cultural computing where robust semantic preservation is essential.