LEKH-Net: A Dual-Branch Neural Architecture for Computational Stylistics in Hindi Poetry

Authorship attribution in low-resource Indic languages remains a challenging frontier, particularly for poetic forms like the Ghazal, where authorial identity is encoded not just in semantic themes but in rigid morphological structures (Beher). While standard transformers (e.g., Hindi-BERT) are proficient in thematic classification (~90%), this study reveals that they struggle to distinguish stylistically similar poets due to the loss of sub-word morphological resolution. To bridge this fine-grained gap, we introduce LEKH-Net, a hybrid Dual-Branch architecture. LEKH-Net synergizes a Transformer-based Semantic Branch for word-level context with a 1D-CNN-based Character-Morphology Branch for local orthographic feature extraction. Extensive benchmarks on a dataset of Hindi Ghazals demonstrate that while Transformers are strong baselines, LEKH-Net achieves a state-of-the-art accuracy of 92.5%. Crucially, ablation studies confirm that the inclusion of the morphological branch provides a statistically significant stabilization (p < 0.001), resolving ambiguities that purely semantic models misclassify