Uncertainty-aware joint modeling for Sanskrit compound splitting and segmentation

Sanskrit compound word splitting is challenging because Sandhi-driven phonological transformations obscure word boundaries, making splitting non-deterministic, especially in multi-split cases where multiple hidden boundaries must be identified and constituent segments reconstructed. In this study, a joint end-to-end Transformer-based multi-task architecture is proposed to address this problem by integrating boundary detection and segmented sequence generation within a unified framework. The model employs a shared character-level Transformer encoder that feeds a BiLSTM-CRF boundary prediction branch, which produces sequence-consistent boundary locations and posterior marginals, and a Transformer decoder that generates the segmented output autoregressively. To couple the two tasks, boundary probabilities and entropy-based uncertainty are derived from the CRF marginals. These signals are used to gate encoder representations and to construct a global boundary-aware context that conditions each decoding step, thereby enabling more robust decoding under boundary ambiguity. Experimental results demonstrate consistent improvements over state-of-the-art methods. The model achieves exact-match boundary location accuracy of 84.39%, exact segmentation accuracy of 79.34%, and character-level accuracy of 87.79%. It also achieves boundary-level Precision, Recall, and F1 scores of 92.84%, 91.66%, and 92.25%, respectively. These results indicate that uncertainty-aware coupling and structured BiLSTM-CRF supervision improve segmentation performance while maintaining strong boundary detection, thereby enabling more accurate morphological analysis for NLP and digital humanities applications.