An explainable hybrid deep learning framework for computational aesthetics, thematic mining, and sentiment analysis...

Poetry presents substantial challenges for natural language processing due to its metaphorical abstraction, emotional ambiguity, symbolic complexity, and irregular linguistic structure. Existing computational literary analysis approaches rely on shallow semantic representations, limiting their ability to capture the multidimensional nature of poetic language. This study proposes an Explainable Hybrid Deep Learning Framework (E-HDLF) for integrated computational aesthetics modeling, thematic mining, and sentiment-emotion analysis in English poetry. The framework combines transformer-based contextual embeddings, Bidirectional Long Short-Term Memory (BiLSTM) sequential modeling, attention mechanisms, handcrafted linguistic descriptors, thematic semantic vectors, and explainable artificial intelligence techniques within a unified multi-task architecture. A curated corpus of 12,480 English-language poems was analyzed using RoBERTa and DeBERTa contextual embeddings. Comparative evaluation against SVM, CNN, BiLSTM, and transformer-only models demonstrated superior performance of the proposed framework, achieving thematic classification accuracy of 93.8%, emotion classification F1-score of 0.92, and computational aesthetics regression R² of 0.84. Explainability analyses using SHAP, LIME, attention heatmaps, and fuzzy-rule inference confirmed that predictions were driven primarily by semantically meaningful contextual and figurative structures. Although performance decreased for highly abstract and culturally dependent metaphors, the findings demonstrate that explainable hybrid transformer architectures provide an effective and interpretable framework for advanced computational literary intelligence and poetic text analysis.