Darshana Graph: A Parallel Commentary Corpus for Comparative Indian Philosophy, with Stylometric and Exploratory Graph Analyses

Technical Summary: Darshana Graph

Problem Statement
Classical Indian philosophy is fundamentally characterized by centuries of argumentation among distinct traditions (Hindu, Buddhist, and Jain) interpreting the same foundational texts. However, existing digital resources largely fail to capture this comparative dimension. Most available datasets present single translations with a single point of view, or aggregate multiple texts without structuring them to allow direct comparison of how different commentators interpret the same root verse or sutra. Consequently, machine-readable resources lack the structural alignment necessary to study interpretive divergence across schools.

Methodology
The authors constructed Darshana Graph, a corpus of over 125,000 text records drawn exclusively from public-domain and openly licensed sources. The corpus is dominated by the Pali Canon (114,591 records) but includes a structurally unique subset of approximately 8,500 Hindu and Jain records.

• Corpus Construction: The core innovation is the alignment schema. A single record represents a root verse or sutra (e.g., Brahma Sutra 1.1.1) and contains parallel fields for multiple commentaries from different schools (e.g., Advaita, Vishishtadvaita, Dvaita). The authors addressed significant data quality issues during construction, specifically identifier collisions caused by scraping functions that reset internal counters, which required regenerating globally unique identifiers to prevent silent data loss.
• Stylometric Analysis: The first analysis employs purely statistical, non-machine learning methods to compare argumentative styles. The authors computed five metrics: average passage length, words per sentence, vocabulary diversity, the proportion of passages with explicit scriptural citations, and the proportion with explicit refutation markers. These markers were identified via simple, manually defined regular expressions.
• Knowledge Graph Extraction: The second analysis utilizes a constrained Large Language Model (LLM) pipeline (Llama 3.1 8B) to extract typed philosophical relationships. To ensure auditability, the model is restricted to a closed, predefined vocabulary of thirteen relation types and specific school labels. Any output falling outside this vocabulary is mechanically discarded post-generation, regardless of syntactic validity. This approach prioritizes auditability over recall.

Key Results

• Stylometric Patterns:
  • Citation vs. Refutation: A moderate negative correlation was observed between scriptural citation density and explicit refutation rates across commentators. Those relying heavily on accumulated textual proof (e.g., Madhva) tended to spend less effort explicitly naming and refuting rival positions, while those with lower citation rates (e.g., Nimbarka, Srinivasa) exhibited higher refutation rates.
  • Generational Trends: Within a specific doctrinal lineage, a striking increase in refutation rate was observed across three figures spanning roughly six centuries (Madhva at 2.0%, Nimbarka at 13.2%, and Srinivasa at 42.0%). The authors note this as a pattern worthy of specialist attention rather than a statistically validated trend, citing confounds such as the specific genre of Srinivasa's sub-commentary.
  • Pali Canon Structure: Internal analysis of the Pali Canon revealed measurable genre-level differences in passage length, with aphoristic texts like the Dhammapada averaging 31 characters per segment versus more discursive collections like the Udana at 74 characters.
• Concept Graph Findings:
  • The extraction pipeline successfully identified cross-school disagreement patterns. The most contested pair, atman and brahman, showed Advaita asserting identity (IS_IDENTICAL_TO) while Vishishtadvaita and Dvaita asserted distinctness (IS_DISTINCT_FROM), grounded in different evidence passages.
  • A second contested pair, atman and jiva, showed Advaita asserting distinctness while other schools asserted identity, reflecting standard doctrinal treatments of the empirically conditioned self.
• Limitations and Failure Modes:
  • Approximately 70% of extracted edges carried a general rather than specific school attribution, attributed to model limitations rather than source ambiguity.
  • The relation type IS_QUALIFIED_ASPECT_OF was significantly over-represented, suggesting the model defaulted to this label when uncertain.
  • An embedding-based, extraction-free method used to corroborate findings produced results in direct tension with the LLM findings for the atman-brahman pair, highlighting a limitation of topic-level sentence embeddings in capturing propositional philosophical agreement.

Significance and Claims
The paper positions the alignment structure of the corpus—linking the same source material to eighteen distinct commentators across five schools—as its primary and most defensible contribution, noting that such a structure is not available elsewhere at this scale.

The authors explicitly frame the stylometric analysis as a transparent, reproducible starting point for philological reading rather than a validated theory of rhetorical strategy. Similarly, the knowledge graph extraction is presented as "genuinely useful but explicitly less mature," with a candid account of its failure modes. The significance of the work lies in:

1. Providing a resource that enables the direct computational comparison of independent interpretive traditions.
2. Demonstrating a constrained LLM pipeline that trades recall for auditability in historical text analysis.
3. Offering a documented case study on the practical difficulties of building and honestly evaluating LLM-assisted knowledge extraction from multi-perspective textual traditions.

The authors release the full corpus, the extracted relationship graph, and all construction code, inviting further work on formal precision evaluation and the expansion of the corpus to address identified gaps in school coverage and sub-tradition distinction.