Towards Traceability from LLMs: Leveraging Sanskrit For Epistemological Analysis.

Currently, widely available Large Language Models (LLMs) cannot ‘explain’ their outputs to humans, fundamentally due to their underpinnings in deep learning (DL) technology. In this paper we investigate how the unusual properties of the ancient Sanskrit language could offer both semantic and ontological traceability in the context of LLMs. We show how the formal definitions of Sanskrit grammar and the Nyāya-Vaiśeṣika (N-K) philosophical system could provide informationally strong and rigorous traceability of LLM outputs, together with associated epistemological analysis including: Semantics within an infinitely-extensible lexicon that is rooted in an established finite set of stems; and Ontological commitments, within a particular type of philosophical realism. These capabilities transcend those of Web Ontology Language (OWL) in providing: A rich set of types of relationship through inherence; A form of negation that transcends modern logic; and Explicit support for distinguishing types of causality.