Predictive Quality Management in Nepalese Deciphering Ethics through Language: A Quantitative and Exploratory Study on Sanskrit Semantics and Generative AI for Ethical Intelligence

Abstract

Generative Artificial Intelligence, particularly Large Language Models (LLMs) based on transformer architectures, has demonstrated remarkable linguistic and reasoning capabilities. However, its effectiveness in modeling deep ethical semantics remains underexplored, especially in the context of classical languages. This study investigates the suitability of LLMs for interpreting Sanskrit—a language renowned for its precision, semantic depth, and philosophical richness—for cultivating ethically grounded AI systems. Three pilot datasets of 300 items each were constructed for semantic similarity, ethical inference, and contextual reasoning, drawing on the Bhagavad Gita and related ethical vocabulary. A multilingual sentence-transformer baseline was evaluated on the semantic similarity task against two structurally enhanced configurations — a character/word-overlap enhancement and a Heritage.py-derived morphological enhancement — using Pearson and Spearman correlation and mean absolute error. Contrary to the initial hypothesis, neither enhancement improved on the transformer baseline: the character/word signal collapsed to zero due to a script mismatch between the paired Sanskrit strings, and the morphological signal produced a small but statistically significant negative correlation with the pilot reference scale (Pearson r = −0.18, p = 0.002). A category-level diagnostic further revealed that the baseline transformer does not reliably separate semantically equivalent from semantically unrelated Sanskrit expressions on this pilot set. The ethical inference and contextual reasoning datasets (300 items each) and their evaluation pipelines have been fully constructed and implemented but had not completed model evaluation and expert annotation review at the time of this submission. The paper reports these findings transparently, diagnoses their concrete causes, and proposes a five-layer hybrid ethical intelligence architecture together with a corrective roadmap for the next phase of evaluation.