📄 Quantifying Text Reuse Across Three Kṛṣṇa Yajurveda Recensions: Using Multi-Algorithm Computational Collation

Summary

The Kṛṣṇa Yajurveda survives in parallel recensions whose textual overlap had never been measured systematically. The authors compare the Maitrāyaṇī Saṃhitā (MS), Kāṭhaka Saṃhitā (KS), and Taittirīya Saṃhitā (TS) in two ritual sections: Agnyupasthāna, the daily worship of the fires, and Punarādhāna, the re-establishment of the fires. Each section was prepared in a plain romanized form and a lemmatized form. For the analysis they built ICoMa (Intertextuality Collation Machine), a free browser-based tool that runs several similarity algorithms on the same data so that results can be checked against each other. Every independent algorithm ranks the pairs the same way: MS–KS closest, then KS–TS, then MS–TS. The two sections behave very differently. In Punarādhāna, MS and KS are nearly identical and follow the same order, while TS diverges, which points to direct borrowing. Agnyupasthāna shows moderate similarity spread across all three pairs. The authors read this as evidence that sections were transmitted along different paths and were fixed as texts at different times. The study also shows that character-level comparison of plain text suits these orally preserved sources well.

Key points

• The tool: ICoMa is written in TypeScript/React and reportedly supports ten script systems. It implements six measures: Levenshtein, character n-gram, word n-gram (Jaccard), character Jaccard, Smith–Waterman, and script-aware normalization. The character n-gram measure gave results identical to Levenshtein at window size n = 3, so five measures are independent. The tool and the data are openly released.
• Punarādhāna: MS–KS coverage reached 93.5% (Levenshtein, plain text; 85% mean similarity). MS–TS coverage was only about 20%.
• Agnyupasthāna: Coverage was 47–55% for MS–KS, 30–38% for KS–TS, and 27–34% for MS–TS.
• Negative control: An unrelated MS chapter (Darśapūrṇamāsa-Mantras) scored only 1.9–3.1% coverage. The exception was Jaccard, which scored 5.7% on lemmatized and 13.0% on plain text, so the authors judge Jaccard alone too undiscriminating for this task.
• Data and setup: The lemmatized texts were supplied by Oliver Hellwig, following Nehrdich et al. (2024). Thresholds were calibrated so that MS–KS Agnyupasthāna (lemmatized) scores about 50%.
• Stated limitations: Thresholds were set by inspecting histograms by hand, the control was applied to only one section, and the study covers just two sections.