Mapping Hymns and Organizing Concepts in the Rigveda: Quantitatively Connecting the Vedic Suktas

Venkatesh Bollineni, Igor Crk, Eren GultepeNAACL 2025

TL;DR

This paper tackles the challenge of understanding the ancient Rigveda by using Natural Language Processing (NLP) to map its hymns (suktas) and uncover thematic connections. They introduce a novel "mean-LSA" embedding technique that, unlike deep learning methods, successfully identifies statistically significant thematic clusters within the Rigveda, aligning with known groupings.

What Problem Does This Paper Address?

The Rigveda, a foundational text of Hinduism, is incredibly ancient and complex, making it difficult for both scholars and laypeople to navigate and grasp its concepts. Unlike more modern religious texts with extensive commentaries and accessible resources, the Rigveda's sheer volume, poetic structure, and ancient Sanskrit language present significant barriers. Existing NLP research often focuses on more recent or Abrahamic religions, leaving a gap in computational analysis of Hindu scriptures. This paper aims to bridge this gap by quantitatively organizing the Rigveda's hymns and revealing their underlying thematic structure without relying on prior human-defined knowledge.

Methods

The core of this research lies in creating numerical representations (embeddings) of each hymn (sukta) and then analyzing the relationships between these embeddings to form networks. They explored three main embedding techniques:

  • mean-LSA: This is the paper's novel contribution. It's an adaptation of Latent Semantic Analysis (LSA). Instead of chunking documents into fixed sizes (which can lose semantic information), mean-LSA takes the average of word embeddings within a sukta to create a single, representative embedding for that sukta. This method aims to preserve the original sukta's semantic content.
  • SBERT (Sentence-BERT): A pre-trained transformer model designed for sentence embeddings, known for its effectiveness in clustering and similarity tasks.
  • Doc2Vec: A technique that creates document embeddings by learning to predict words within a document.

After generating these embeddings, they used UMAP (Uniform Manifold Approximation and Projection) for dimensionality reduction to make the data computationally tractable and visually interpretable. Then, they constructed "sukta networks" by identifying the k-nearest neighbors for each sukta embedding. Finally, community detection algorithms (Louvain, Leiden, and label propagation) were applied to these networks to identify clusters of suktas that likely represent thematic topics. A crucial innovation was their method for testing the statistical significance of these detected communities by permuting the network adjacency matrices, ensuring the identified structures weren't due to random chance.

Datasets

The primary dataset used is the modern English translation of the Rigveda by Jamison and Brereton (2014). This translation comprises 10 books (mandalas), 1,028 hymns (suktas), and 10,552 verses (mantras). No new benchmark datasets were introduced; the focus was on analyzing an existing, well-established text.

Key Insights and Ideas

  • The power of "mean-LSA": The paper's novel mean-LSA embedding method significantly outperformed deep learning-based methods (SBERT and Doc2Vec) in identifying statistically significant and semantically meaningful thematic clusters within the Rigveda. This suggests that simpler, adapted traditional methods can still be highly effective, especially when tailored to the specific characteristics of the data.
  • Statistical significance is crucial: Simply achieving a high modularity score in community detection doesn't guarantee meaningful results. The rigorous statistical significance testing, using random permutations, was essential in distinguishing genuine thematic structures from random noise. This highlights the importance of robust validation in unsupervised learning tasks.
  • Deep learning methods struggled: Despite their sophistication, SBERT and Doc2Vec embeddings failed to produce statistically significant networks. SBERT even mistakenly grouped distinct thematic suktas together, indicating a lack of fine-grained semantic discrimination for this specific task. Doc2Vec was the least successful, failing to detect meaningful clusters.
  • Alignment with known groupings: The mean-LSA method successfully identified clusters that corresponded to seven well-known thematic groupings of Rigvedic hymns (e.g., Creation, Funeral, Water), validating its ability to uncover historically recognized conceptual divisions.

Key Results

  • The mean-LSA embedding method produced a statistically significant network (z = 2.726, p < .01) with a high modularity score of 0.944.
  • This mean-LSA network was successful in identifying clusters for all seven famous sukta groupings analyzed.
  • SBERT produced a network with a high modularity (0.950) but was not statistically significant (z = -0.876, p = .810). It correctly identified 62.7% of famous suktas but mistakenly combined three distinct topics into a single "Mixed" group.
  • Doc2Vec also had a high modularity score (0.952) but was not statistically significant (z = -0.126, p = .550) and failed to detect relevant suktas, especially for the seven famous cases.
  • The mean-LSA method correctly identified 71.9% of famous suktas across the seven cases.

Limitations and Future Work

  • Translation dependency: A primary limitation is the reliance on a single modern English translation. The authors acknowledge that translations can mask the original meaning of the Vedic Sanskrit, and future work should compare results with the original Sanskrit version.
  • Future directions: The authors plan to investigate the training of unsupervised transformer language models and explore the application of their statistical significance testing method to other NLP tasks, such as coherence measures in topic modeling. They also suggest comparing the Rigveda with its Sanskrit version for future studies.

Summary updated

BibTeX
@inproceedings{bollineni2025mapping,
  author = {Venkatesh Bollineni and Igor Crk and Eren Gultepe},
  title = {{Mapping Hymns and Organizing Concepts in the Rigveda: Quantitatively Connecting the Vedic Suktas}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure