AI-Based Hindu Shloka Understanding and Context-Aware Meaning Generation Using Natural Language Processing and Large Language Models |
Author(s): |
| Pooja Yagnik , Gandhinagar University |
Keywords: |
| Sanskrit NLP, Shloka Interpretation, Large Language Models, Retrieval-Augmented Generation, Machine Translation, Low-Resource Languages, Hallucination, Indian Knowledge Systems |
Abstract |
|
Sanskrit shlokas (verses) are compact, metrically constrained texts whose meaning depends on word-level morphology, compound structure, the work in which they occur, the speaker and situation, and the interpretive tradition of the reader. General-purpose large language models (LLMs) can paraphrase well-known verses, yet they frequently flatten polysemy, ignore commentarial context, and produce fluent but unfaithful explanations. This paper presents the design of a hybrid, context-aware system that takes a Sanskrit shloka as input and produces word-level glosses, a literal translation, and an explanatory contextual meaning in English and Hindi. The architecture combines (i) Devanagari/IAST normalisation and metre identification, (ii) a linguistic analysis layer for sandhi splitting, compound analysis and morphological tagging, (iii) retrieval-augmented generation (RAG) over an indexed corpus of verses and traditional commentaries, and (iv) a prompted or parameter-efficiently fine-tuned LLM with source-grounded output constraints. An evaluation protocol is specified that pairs automatic metrics (BLEU, chrF, BERTScore, COMET) with expert human assessment of faithfulness, fluency, contextual appropriateness and hallucination, together with ablations that isolate the contribution of each component. The paper contributes a reproducible methodology, an error taxonomy for shloka interpretation, and a discussion of the responsible handling of sacred texts, in which multiple legitimate interpretations are surfaced rather than collapsed into a single answer. |
Other Details |
|
Paper ID: IJSRDV14I90055 Published in: Volume : 14, Issue : 9 Publication Date: 01/12/2026 Page(s): 107-110 |
Article Preview |
|
|
|
|
