Sanskrit Word Sense Disambiguation Based on Lexicographic Definitions

We present a word sense disambiguation (WSD) system for Vedic and Classical Sanskrit that extends the gloss reader architecture of Blevins and Zettlemoyer (2020). Our work addresses two key questions: how semantic inventory choice affects WSD accuracy, and which data augmentation strategies can improve performance. Using the Sanskrit-English dictionary of Monier-Williams as a test case, we explore whether traditional lexicographic resources with overlapping and periphrastic definitions can be used to train a WSD system. Our model achieves high accuracy on dominant senses but struggles with secondary senses, a pattern confirmed by statistical error analysis. While model architecture has limited impact, the granularity of the semantic inventory substantially affects performance. For augmentation, we identify alignment-based and LLMbased methods as promising approaches, particularly for increasing coverage of underrepresented senses across genres and time periods.