In Natural Language Processing (NLP), resolving lexical and semantic ambiguity remains a formidable challenge. This is particularly true for Sanskrit, where a single sentence can often be morphologically parsed in multiple valid ways. For instance, in the sentence Yānaṃ vanaṃ gacchati ("The vehicle goes to the forest"), a machine might struggle to determine whether the "vehicle" or the "forest" is the agent of the action. While humans use common sense to dismiss the idea of a forest "going," machines lack this inherent semantic grounding.
This paper introduces a computational approach to Word Sense Disambiguation (WSD) for Sanskrit by formalizing the concept of Yogyatā—semantic compatibility between words. By leveraging ancient Indian linguistic and philosophical theories, the authors develop a framework to help machines automatically prune incorrect semantic interpretations.
Methodology
The authors propose a system based on Kāraka-Yogyatā, which defines the semantic relatedness and expectancy between a verbal root (dhātu) and a noun. The methodology involves three core components:
- Ontological Tag-set Extension: Building upon existing Sanskrit ontologies to categorize words by functional capabilities (e.g., distinguishing an "instrument of motion" from a generic "inanimate object").
- Annotation Tool Development: A custom web-based PHP/MySQL interface was created to allow linguists to manually link verbal roots (L-words) with dictionary lexemes (R-words).
- Semantic Constraint Mapping: Annotators define valid kāraka (thematic roles like agent, object, or locus) for specific word senses, creating a database for structural constraints.
Key Insights and Ideas
A central insight of this work is that ancient Paninian grammar and the philosophical schools of Nyāya and Vyākaraṇa provide highly structured frameworks that map well to modern computational constraint satisfaction problems. Yogyatā acts as a semantic filter, shifting WSD from statistical guessing to logical pruning.
The importance of contextual ontology is highlighted by words with multiple meanings. For example, the word kaṃsa can refer to a "metal vessel" or the "King of Mathura." Its valid grammatical roles change depending on the sense used and the verb it interacts with.
As shown in the table above, kaṃsa can only act as an apādāna (point of separation/source) for certain verbs if it is used in the "King of Mathura" sense. This demonstrates that static part-of-speech tagging is insufficient for deep parsing; semantic sense must be integrated with grammatical expectancy.
Results and Case Study
The authors successfully deployed the Yogyatā Relation Tool and validated their approach through a case study on the apādāna relation. By mapping specific verbal roots (such as añcu, ñibhī, and pā) against various nouns, they demonstrated how the system distinguishes between ambiguous word senses based on Panini's aphorisms (1.4.24 - 1.4.31).
Limitations and Future Work
While promising, the methodology currently relies on time-consuming manual annotation by experts, making it difficult to scale rapidly. Furthermore, quantitative benchmarks are needed to measure the exact improvement in parser accuracy. Future work will focus on scaling the database, applying the framework to Machine Translation, and exploring cognitive psycholinguistics to verify the necessity of Yogyatā in human verbal cognition.
Summary updated