💬 NLP

Three Buddhist Vocabularies: Computational Stylometry of the English Pali Canon across Sutta, Vinaya, and Abhidhamma

This paper presents a computational stylometric analysis of the English Pali Canon across the Sutta, Vinaya, and Abhidhamma Pitakas, revealing distinct lexical profiles for each corpus—such as the Abhidhamma's higher diversity and numeral density, the Theravada and Mulasarvastivada Vinayas' shared legal heritage, and significant translation variations over time—while confirming that all texts adhere to Zipf's law.

Original authors: Joy Bose

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joy Bose

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Buddhist Canon (the Tipitaka) as a massive library containing three distinct wings, each written by different authors for different purposes. For a long time, researchers have studied the "Sutta" wing (the stories and sermons). This paper is like a new librarian who decides to walk into the other two wings—the Vinaya (the rulebook for monks) and the Abhidhamma (a technical encyclopedia of the mind)—and run a computer program to see how the language itself changes from room to room.

The researcher, Joy Bose, didn't read every word to understand the deep philosophy. Instead, they used a "word counter" and a "pattern detector" to see how the vocabulary behaves. Here is what they found, explained simply:

1. The Three "Flavors" of Language

Think of the three sections as three different types of parties:

The Sutta Wing (The Storytellers): This is the biggest room. The language here is like a campfire story. It repeats the same phrases often (like "the monk said" or "four things"). The computer found that the words here are very repetitive, creating a steady, rhythmic beat.
The Vinaya Wing (The Rulebook): This room is full of legal instructions. It's surprisingly similar to the Storytellers in how repetitive it is. Whether you read the modern translation (Brahmali) or the old one (Horner), the "vocabulary rhythm" is almost identical. It's like two different people reading the same phone book; the words might change slightly, but the structure is the same.
The Abhidhamma Wing (The Encyclopedia): This is the odd one out. It's a technical manual listing 52 types of feelings and 89 types of consciousness. Because it has to name so many specific, unique things, it uses a much wider variety of words. It's less repetitive and feels more like a modern science textbook than a story.

2. The "Zipf" Test: Who is the Star of the Show?

The researchers used a famous math rule called Zipf's Law. Imagine a rock concert where one singer sings 50% of the songs, the second singer sings 25%, and the rest split the remaining time. In most texts, a few common words (like "the," "and," "it") dominate the top spots.

In the Stories and Rules: The top 10 words are mostly grammar helpers (like "it" or "they").
In the Encyclopedia: The computer found something weird. The word "Consciousness" jumped into the top 10, pushing the grammar words down. It's as if, in a library of technical manuals, the word "Engine" appears so often it drowns out the word "The." This proves the Encyclopedia is obsessed with specific concepts, not just storytelling.

3. The "Vocabulary Overlap" (How much do they share?)

The researchers asked: If you take a handful of words from the Rulebook, how many of them would you also find in the Storybook?

The Result: About half of the words in the Rulebook are also in the Storybook. They share a common language because they are part of the same tradition.
The Time Travel Test: The researchers compared the Theravada Rulebook (used in Sri Lanka/Thailand) with the Mulasarvastivada Rulebook (used in Tibet) from 2,000 years ago. Even though they split up two millennia ago, they still share 50% of their vocabulary. It's like finding that a modern American English dictionary and a 19th-century British one still share half their words, proving they came from the same family tree.

4. The "Translation Shift" (How language changes over time)

The researchers took the exact same Rulebook text and compared two translations: one from 1938 and one from 2026.

The Shock: They only shared 24% of their vocabulary! That's a huge gap for the same text.
The Cause: It wasn't that the meaning changed, but the words changed.
- Example 1: In 1938, a deep meditation state was called "musing" (which sounds a bit like daydreaming). In 2026, it's called "absorption" (which sounds much more focused).
- Example 2: A serious offense was called "defeat" in 1938, but "expulsion" in 2026.
- Example 3: The 2026 translator was more direct about anatomy and gender issues, while the 1938 translator used polite, vague language.

5. The "Number" Count

The Encyclopedia (Abhidhamma) is obsessed with counting. It uses number words (one, two, three, four) 3.26% of the time. The Storybook uses them only 2.09% of the time. This makes sense: if you are listing 89 types of consciousness, you have to say "one," "two," "three" a lot more than if you are just telling a story about a monk walking in the woods.

Summary

This paper is like a linguistic X-ray. It doesn't tell us what the Buddha taught in deep detail, but it shows us how the teaching was packaged.

Stories are rhythmic and repetitive.
Rules are procedural and shared across time.
Encyclopedias are dense, technical, and full of unique terms.
Translations change drastically over 88 years, swapping old-fashioned words for modern, precise ones.

The researcher concludes that while we can see these patterns clearly with computers, the deeper question of how these ideas evolved over time requires a more human, concept-level analysis, which is the next step in their work.

Technical Summary: Computational Stylometry of the English Pali Canon

Problem Statement
While computational analysis of Buddhist texts has begun to emerge, previous studies have largely been limited to specific subsets of the Tipitaka (e.g., the Sutta Pitaka in Pali or keyword extraction within single Nikayas). There is a lack of unified, cross-corpus computational analysis covering all three major divisions of the Pali Canon (Sutta, Vinaya, and Abhidhamma) in English translation. Furthermore, existing work has not systematically compared the lexical structures of different Vinaya traditions (Theravada vs. Mulasarvastivada) or analyzed translation shifts across nearly a century of English renderings of the same source text. This paper addresses these gaps by presenting a descriptive, lexical-level analysis of the entire English Tipitaka corpus.

Methodology
The study constructs a corpus of approximately 134,831 text segments from five primary sources:

Sutta Pitaka: 114,591 sentence-level segments from Bhikkhu Sujato's CC0 translation (bilara-data).
Vinaya Pitaka: 7,923 paragraph-level segments from Bhikkhu Brahmali's 2026 translation; 2,826 segments from I.B. Horner's 1938 translation; 7,229 segments from the 84000 Mulasarvastivada translation; and a preliminary 185-segment Dharmaguptaka corpus.
Abhidhamma-adjacent: 2,077 segments from three English translations of the Abhidhammattha Sangaha (a 11th-century compendium), serving as a proxy for Abhidhamma vocabulary.

The analysis employs four primary computational metrics:

Zipf Rank-Frequency Analysis: OLS-fitted power-law exponents on the top 200 words to measure deviation from the ideal Zipf slope (-1).
Lexical Diversity (MATTR-500): Moving Average Type-Token Ratio with a 500-word window to control for corpus size and segmentation differences (sentence vs. paragraph).
Numeral Density: Percentage of tokens matching a 32-term list of numerals.
Vocabulary Overlap: Jaccard similarity and the Szymkiewicz-Simpson overlap coefficient to measure shared vocabulary between corpora, with the latter used to mitigate bias from corpus size imbalances.

All analyses are descriptive point estimates; no formal significance testing or hypothesis testing is conducted.

Key Results

Zipf Structure: All four major corpora exhibit strong Zipfian distributions ( $R^2 > 0.989$ ). The Theravada Vinaya is closest to the ideal slope (-0.949), while the Sangaha corpus deviates most (-0.876). This deviation is characterized by a "flatter head," where technical terms like "consciousness" displace grammatical particles in the top 10 ranks (appearing at rank 8), a phenomenon not seen in the Sutta or Vinaya.
Lexical Diversity: Raw TTR suggested the Sutta was less diverse than the Vinaya, but this was attributed to segmentation artifacts (sentence-level vs. paragraph-level). When measured via MATTR-500, the Sutta (0.399) and Theravada Vinaya (0.400) are nearly identical. In contrast, the Sangaha corpus is genuinely more diverse (0.560), confirmed by size-controlled subsampling. This high diversity reflects the compendium's format of defining distinct categories rather than the formulaic repetition found in the Suttas.
Numeral Density: The Sangaha corpus has the highest density of numeral words (3.26%), consistent with its systematic enumeration of mental and material categories. The Mulasarvastivada Vinaya follows (2.76%), while the Sutta (2.09%) and Horner's 1938 Vinaya (1.62%) are lower.
Cross-Tradition Overlap: The Theravada and Mulasarvastivada Vinayas share 20.0% vocabulary (Jaccard) and 49.1% (overlap coefficient), indicating substantial shared legal heritage despite two millennia of separate development. The Dharmaguptaka corpus shows a low Jaccard (0.082) due to its small size, but a high overlap coefficient (0.816), suggesting most of its limited vocabulary is contained within the Theravada tradition.
Translation Shift: Comparing Brahmali (2026) and Horner (1938) translations of the same Vinaya text reveals only 24.2% vocabulary overlap. Diagnostic shifts include "musing" (Horner) vs. "absorption" (Brahmali) for jhana, and "defeat" vs. "expulsion" for parajika. Approximately 20% of this divergence is attributed to spelling variations and omitted sections rather than terminological shifts.

Significance and Claims
The paper claims to extend earlier computational findings (specifically Zigmond, 2020) from the Pali original to English translations, demonstrating that the three Pitakas remain computationally distinct even in translation. The study attributes the distinctiveness of the Abhidhamma-adjacent corpus not to the canonical Abhidhamma Pitaka itself, but to the specific "compendium format" of the Abhidhammattha Sangaha, which generates high lexical diversity through enumeration rather than repetition.

The authors position this work as a descriptive foundation for future research. The paper explicitly states that while it documents structural and lexical differences, it does not answer deeper doctrinal questions (e.g., the evolution of Vinaya rules or the absorption of Abhidhamma categories into Mahayana). Instead, it frames these results as the necessary lexical baseline for future concept-level analysis using the Darshana Graph (Bose, 2026a), specifically by extending IS_PRECURSOR_OF and NEGATES edges to include these new corpora. The paper emphasizes that all code and data are released as open-source extensions to the Darshana Graph dataset.