Ancient medical traditions represent centuries of accumulated clinical observation, pharmacological knowledge, and surgical technique. The Sushruta Samhita, one of the foundational texts of Ayurvedic medicine, contains detailed descriptions of diseases, herbal treatments, and surgical procedures that predate much of Western medical codification. The problem is access. These texts exist in archaic Sanskrit, with translated versions that are themselves dense and domain-specific, making systematic computational analysis genuinely difficult. This paper, available at arXiv:2608.28608, applies a pipeline of NLP methods to translated versions of these texts, aiming to extract structured knowledge and make it queryable.
The work sits at an interesting intersection of digital humanities, medical informatics, and information retrieval. It is not the first attempt to apply NLP to historical or traditional medical corpora, but the specific combination of tools and the target domain give it some genuine novelty.
Methodology
The authors construct a three-stage pipeline. First, Named Entity Recognition (NER) is applied to identify and classify entities across several categories: diseases, treatments, medicinal plants, and researchers. Second, BERTopic modelling is used to identify latent thematic structure across the corpus, grouping passages by their dominant medical subject matter. Third, the extracted entities and their relationships are loaded into a Neo4j graph database, enabling semantic querying and visualisation of how concepts relate to one another across the text.
The choice of BERTopic is sensible here. Unlike LDA-based approaches, BERTopic uses sentence-level transformer embeddings clustered via UMAP and HDBSCAN, which tends to produce more coherent and interpretable topics on shorter, domain-specific text segments. For a corpus with specialised vocabulary and relatively short passages, this is a reasonable methodological choice. The NER component would have required domain-specific training or fine-tuning, and the paper would benefit from being clearer about exactly how the entity recognition model was adapted for Ayurvedic terminology, which does not appear in standard biomedical NER training sets like NCBI-Disease or BC5CDR.
The Neo4j knowledge graph is the most practically useful output. Graph-based representations of medical knowledge have a well-established history in biomedical informatics, and encoding relationships between, say, a specific plant, the condition it treats, and the preparation method described in the text allows for queries that flat document retrieval simply cannot support.
Key Contributions
- A working NLP pipeline applied to translated Ayurvedic texts, combining topic modelling, entity recognition, and graph-based storage
- Thematic classification of medical content using BERTopic, producing interpretable topic clusters across the Sushruta Samhita
- A Neo4j knowledge graph encoding relationships between diseases, treatments, plants, and procedures extracted from the corpus
- A demonstration that computational methods can surface structure in historical medical texts that is not obvious from linear reading
Results and Findings
The paper reports that BERTopic successfully identified coherent medical themes across the corpus, with topic clusters corresponding to recognisable Ayurvedic domains such as wound treatment, herbal pharmacology, and anatomical description. The NER pipeline extracted entities across the defined categories with reasonable coverage, though the paper does not report precision and recall figures against a held-out annotated test set, which is a significant omission for any NLP evaluation.
The knowledge graph visualisations show meaningful clustering of related concepts. Medicinal plants appear connected to multiple disease entities, treatments link to both conditions and plant-based preparations, and the graph structure reveals relationships that would be tedious to extract manually at scale. The authors argue, plausibly, that this kind of structured representation supports both historical scholarship and potential integration with modern medical informatics systems.
Nineteen pages, eight figures, and five tables give this a reasonable level of documentation for a conference paper. It received a Best Paper Award at the TISS-IATLIS National Conference in 2025, which speaks to its reception within the library and information science community it was presented to.
Limitations and Open Questions
Several methodological gaps deserve attention. The absence of quantitative NER evaluation is the most pressing. Without precision, recall, and F1 scores against annotated ground truth, it is impossible to assess how reliably the entity extraction is actually working. Domain-specific NER on archaic medical vocabulary is hard, and the paper would be considerably stronger with this evidence.
The reliance on translated texts rather than original Sanskrit also introduces a layer of indirection. Translation choices affect which terms appear, how concepts are rendered, and what semantic relationships are preserved. The authors acknowledge accessibility as a motivation for using translations, but the downstream effect on NLP quality deserves more discussion. Different English translations of the same Sanskrit source text can vary substantially.
BERTopic's performance is sensitive to hyperparameter choices, particularly the minimum cluster size in HDBSCAN and the dimensionality reduction settings in UMAP. The paper does not report ablation studies or sensitivity analyses, so it is unclear how stable the topic structure is across different configurations.
There is also a broader question about how this knowledge graph would be maintained, extended, or integrated with existing biomedical ontologies. Connecting Ayurvedic entities to SNOMED CT, MeSH, or the Gene Ontology would substantially increase the utility of the graph for medical informatics applications, but this is left as future work.
Despite these gaps, the paper makes a genuine contribution to a relatively underserved area. Computational analysis of traditional medical systems is both historically valuable and practically relevant given growing interest in ethnopharmacology and the systematic study of traditional medicine. The pipeline described here is reproducible in principle and points toward a scalable approach for other historical medical corpora. Researchers working on digital humanities projects involving non-Western textual traditions will find the methodology worth examining, even where the evaluation could be more rigorous.
Full paper: arXiv:2608.28608