Design of deep learning models for intelligent classification of Chinese and Western cultural classics
Abstract
In the rapidly evolving domain of intelligent systems, accurate classification of culturally grounded textual data is critical for advancing machine understanding across diverse semantic and epistemological domains. Recent advances in computational humanities have highlighted the challenges of modeling classical texts with deep cultural embeddings, particularly when traditions exhibit stark contrasts in rhetorical structure, philosophical orientation, and thematic encoding. Existing models often rely on surface-level linguistic cues or rigid taxonomical mappings, lacking the capacity to resolve deep semantic divergences or integrate symbolic cultural ontologies. To address these limitations, we propose a deep learning framework, CulCodeNet, which employs a multi-level architecture incorporating semantic disentanglement, graph-based structural reasoning, and cross-cultural projection layers to capture nuanced rhetorical and thematic signals. This architecture is further enhanced by a training strategy termed Contrastive Cultural Fusion, which integrates contrastive alignment, ontology-grounded supervision, and curriculum-aware sampling. Experimental evaluations demonstrate that the proposed framework significantly outperforms traditional classification baselines on tasks involving high-variance cultural corpora, achieving superior generalization and interpretability. By combining symbolic representation, hierarchical modeling, and contrastive cultural alignment, this approach establishes a novel paradigm for the intelligent analysis of heritage texts, contributing to advancements in cross-lingual, interpretable, and culturally grounded computing.