Kumar, Deepu2026-09-232026-09-232026http://nits.ndl.gov.in/handle/123456789/127Sanskrit, one of the oldest Indo-European languages, presents unique challenges for natural language processing due to its exceptionally intricate morphological system governed by Pāṇini’s grammar. The accurate classification of Sanskrit word forms into subanta (सुबन्त, nominal), tiṅanta (ितङन्त, verbal), and avyaya (अव्यय, indeclinable) categories serves as a prerequisite for a wide range of downstream tasks, including pars-ing, translation, morphological analysis, and sandhi resolution. Despite the growing availability of annotated Sanskrit corpora, computational processing of the language remains challenging due to its low-resource status, agglutinative morphology, and the widespread occurrence of sandhi.In this thesis, a hybrid architecture is proposed that integrates a fine-tuned MuRIL transformer with a character-level convolutional neural network (CNN) for suffix feature extraction. Four distinct fusion strategies – concate-nation, gated fusion, multi-head cross-attention, and low-rank bilinear pooling – are systematically investigated for combining contextual transformer representations with explicit morphological suffix features. A comprehensive ablation study comprising 23 model configurations (7 baselines and 16 hybrid variants across 4 suffix lengths and 4 fu-sion methods) is conducted, with each configuration evaluated over 3 random seeds for a total of 69 experimental runs on the Universal Dependencies Sanskrit–Vedic treebank. The best hybrid configuration (CNN-SL10-attention) achieves a macro-F1 score of 97.47±0.01%, outperforming all baselines including standalone MuRIL (97.27±0.09%), XLM-R (97.39±0.05%), mBERT (97.25±0.01%), and ByT5 (95.70±0.12%). The hy-brid model also demonstrates the lowest cross-seed variance (σ = 0.009) among all evaluated models, representing approximately a tenfold improvement in stability over the MuRIL baseline. The ablation study reveals that attention-based fusion with longer suffix windows consistently yields the best performance. However, the narrow macro-F1 range across all 16 hybrid variants (97.34–97.47%) together with a non-significant McNemar test result between the hybrid model and MuRIL (p = 0.576) suggests that pre-trained multilingual transformers may already encode much of the morphological information captured by explicit suffix features. This finding has broader implications for the design ofen