Four Distinct Languages of Yoga Emerge When Algorithms Read Its Texts

A new computational study suggests that this is not just an impression: when machine-learning models were set loose on hundreds of passages drawn from four very different bodies of yoga writing, the texts sorted themselves into four sharply distinct semantic clusters, each anchored to one type of source. The research team, Sabbi Lall and Swayam Bagaria of Harvard Divinity School together with Jinjie Liu of the Institute for Quantitative Social Science, began with a deceptively simple question. Yoga originated in Indian contemplative and mystical traditions and has since become a globalized set of psychophysiological and psychospiritual practices.

Its spread has produced classical scriptures, modern practice manuals, a large biomedical literature on yoga-based interventions, and a scholarly field of yoga studies. But does this multiplication of formats actually lead people to describe yoga in converging ways, or do different communities inhabit genuinely different conceptual worlds? To find out, the authors compiled 200 text extracts from translated classical scriptures, recent biomedical research articles, contemporary yoga practice manuals and academic yoga studies publications.

Each extract was built around a locatable "yoga is…"-style statement together with its surrounding context, a design choice that gave the natural language processing pipeline a consistent unit of analysis: an explicit definition or characterization of yoga in situ. The analytical workflow combined several established text-mining techniques. Crucially, the topics were allowed to emerge inductively from the texts themselves rather than being imposed from predetermined scholarly categories.

The number of topics was selected through a coherence-score sensitivity analysis, a standard robustness check that tests how interpretable different model configurations are. Results were visualized with pyLDAvis and t-SNE, two dimensionality-reduction tools that let researchers inspect how cleanly the topics and documents separate in semantic space. The outcome was strikingly tidy.

The analysis produced four well-separated, stable topics, each dominated by vocabulary with a distinct thematic orientation. The first was interventionist and outcome-focused, the language of clinical trials measuring what yoga does to bodies and health outcomes. The second was experiential and phenomenological, characteristic of practice manuals that describe what yoga feels like from the inside.

The third was genealogical and historical, the register of scholars tracing yoga's textual and institutional lineages. The fourth was soteriological and contemplative, the vocabulary of classical scriptures concerned with liberation and contemplative aims. Statistical testing then confirmed what the visualizations suggested: strong, non-random associations between each topic and a specific corpus, with predicted assignment probabilities exceeding 0.7.

The team also identified the specific weighted terms distinguishing each grammar, giving a fine-grained picture of the vocabulary that separates, say, a randomized controlled trial's framing of yoga from a medieval commentary's framing of it. For the authors, the presence of these stable semantic signatures reveals that each of the four corpora possesses a sufficiently distinctive grammar of yoga. Despite vast heterogeneity within each body of writing, four identifiable semantic patterns frame how yoga's practices and aims are articulated.

Why would such clean boundaries persist in a practice as globally mobile and endlessly hybridized as yoga? The researchers propose that the stability may result from canalizing effects: the institutional location where yoga is practiced appears to shape how its practices and goals get framed. A yoga studio producing teacher-training manuals, a university department producing historical scholarship and a medical journal producing intervention studies each exert their own pressures on language, even as more open-ended borrowing of themes may occur at the level of actual practice.

People may blend techniques freely, but the institutional frames through which those techniques are described remain remarkably compartmentalized. The study is explicitly exploratory, and the authors present it as preliminary empirical validation of a pattern rather than a definitive account. Still, the methodological contribution is substantial: an inductive, computational approach that can map the conceptual structure of a contested, multidisciplinary field without presupposing its categories.

The authors argue that making these grammars visible is a first step toward a dialogically plural, mutually enriching field of yoga studies and practice across its different institutional sites.

Subject of Research: Computational text analytics of yoga discourses across classical, pedagogical, scholarly and scientific corpora

Article Title: Grammars of yoga: an exploratory text analytics approach to classical, pedagogical, scholarly, and scientific corpuses of yoga

Article References: Lall, S., Liu, J., & Bagaria, S. (2026). Grammars of yoga: an exploratory text analytics approach to classical, pedagogical, scholarly, and scientific corpuses of yoga. Frontiers in Psychology, 17, Article 1908435. https://doi.org/10.3389/fpsyg.2026.1908435

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1908435

Keywords: yoga, text analytics, topic modeling, LDA, computational humanities, corpus linguistics, cultural psychology, contemplative studies, biomedical research, semantic analysis, Grammars, exploratory