joyboseroy/bengal-dharma-corpus

Bengal Dharma Corpus

Evolution of Bengali Devotional Language: a multi-tradition corpus spanning Old Bengali, Sanskrit, and modern Bengali across Buddhist, Shakta, and Vaishnava traditions, 8th to 19th century.

Assembled and curated by Joy Bose (joyboseroy), June 2026. Code and analysis: https://github.com/joyboseroy/bengal-dharma-corpus Related dataset: joyboseroy/darshana-graph (arXiv:2606.18222)

---

What this corpus is

This is a curated collection of 75 texts from eight devotional traditions of eastern India, spanning approximately 1,200 years of Bengali and Sanskrit religious literature. It is designed for researchers studying the evolution of Bengali devotional vocabulary, Buddhist-Shakta-Vaishnava syncretism, and historical NLP on low-resource Indo-Aryan languages.

The corpus covers three major religious traditions in their full historical arc:

Buddhist tradition: Vajrayana Sanskrit Tara texts (8th-12th c.) and the Charyapada, the oldest surviving Bengali literature (8th-12th c., Old Bengali).

Shakta tradition: Bridge Tara texts presenting the Buddhist goddess Tara in Shakta framing (12th-16th c., Sanskrit), Kali sahasranamas (14th-18th c., Sanskrit), Durga texts, and Bengali Shakta Padavali by Ramprasad Sen (18th c., Bengali, OCR from 1914 edition).

Vaishnava tradition: Gitagovinda by Jayadeva (12th c., Sanskrit), Srikrishna Kirtan by Badu Chandidas (14th-15th c., Old Bengali), Bhakti Rasamrita Sindhu Bengali translation (16th c.), and Chaitanya Charitamrita (16th c., Bengali).

Modern Bengali devotional: Baul songs by Lalon Fakir (18th-19th c., Bengali), Tagore's Gitanjali (19th c., Bengali).

---

Key findings from computational analysis

Using TF-IDF character n-gram vectorization and cosine similarity across the full corpus, the analysis produces the following cross-tradition similarity matrix (selected pairs):

Tradition Pair	Cosine Similarity	Interpretation
Buddhist Tara → Bridge Tara	0.43	Sanskrit Buddhist → Shakta Tara
Bridge Tara → Shakta Kali	0.54	Strongest Sanskrit link
Buddhist Tara → Shakta Kali	0.44	Direct Buddhist-Kali connection
Buddhist Tara → Vaishnava Sanskrit	0.06	Negative control: near zero
Bridge Tara → Vaishnava Sanskrit	0.06	Negative control: near zero
Vaishnava Sanskrit → Shakta Kali	0.00	Key negative control
Shakta Padavali → Baul	0.40	Bengali Shakta → Baul chain
Charyapada → Baul	0.31	Buddhist vernacular → Baul
Vaishnava Bengali → Baul	0.29	Vaishnava Bengali → Baul
Buddhist Tara → Baul	0.02	Near zero: Sanskrit isolation

The most significant finding is the specificity of the Buddhist-Shakta connection. The Bridge Tara tradition is 8.5x more similar to Shakta Kali (0.54) than to Vaishnava Sanskrit (0.06). Vaishnava Sanskrit has near-zero similarity to Shakta Kali. This means the Buddhist-Kali vocabulary overlap is not a generic property of Sanskrit devotional literature but is specific to the Buddhist-Shakta transmission chain.

Three specific Tara texts from the Brihannilatantra tradition show Shakta vocabulary outnumbering Buddhist vocabulary at ratios of 2 to 4 to 1, constituting the strongest evidence for textual migration from Buddhist to Shakta framing.

Vocabulary density gradient across the Sanskrit chain:

Tradition	Buddhist terms/1k	Shakta terms/1k	Ratio
Buddhist Tara	4.2	1.2	0.3 (Buddhist dominant)
Bridge Tara	3.6	3.2	0.9 (near parity)
Shakta Kali	3.8	5.8	1.5 (Shakta dominant)

---

Dataset structure

Table 1: documents

One row per text document.

Column	Type	Description
doc_id	string	Unique identifier
title	string	Full text title
tradition_layer	string	Buddhist_Tara / Charyapada / Bridge_Tara / Shakta_Kali / Shakta_Durga / Shakta_Padavali / Vaishnava_Sanskrit / Vaishnava_Bengali / Baul_Bengali
century_range	string	Approximate composition period
language	string	Sanskrit / Old_Bengali / Bengali / Maithili / IAST
script	string	Devanagari / Bengali / Latin
source_tradition	string	Lineage or source text
source_url	string	Where the text was obtained
n_tokens	int	Token count
buddhist_density	float	Buddhist terms per 1,000 tokens
shakta_density	float	Shakta terms per 1,000 tokens
sb_ratio	float	Shakta/Buddhist ratio
text	string	Full text content

Table 2: pairwise_similarities

One row per document pair, derived from TF-IDF char n-gram cosine similarity.

Column	Type	Description
doc_id_a	string	First document
doc_id_b	string	Second document
tradition_a	string	Tradition of first document
tradition_b	string	Tradition of second document
cosine_sim	float	Cosine similarity (char 3-5 gram TF-IDF)
script_match	bool	Whether both documents use the same script

---

Tradition layers and time periods

8th-12th c.  ── Buddhist_Tara (Sanskrit Vajrayana)
             ── Charyapada (Old Bengali Buddhist)
12th c.      ── Vaishnava_Sanskrit (Gitagovinda, Sanskrit)
12th-16th c. ── Bridge_Tara (Sanskrit, Buddhist-Shakta transition)
14th-15th c. ── Vaishnava_Bengali (Srikrishna Kirtan, Old Bengali)
14th-18th c. ── Shakta_Kali (Sanskrit sahasranamas)
16th c.      ── Vaishnava_Bengali (Chaitanya Charitamrita, Bengali)
             ── Vaishnava_Bengali (Bhakti Rasamrita Sindhu, Bengali)
18th c.      ── Shakta_Padavali (Ramprasad Sen, Bengali, OCR)
18th-19th c. ── Baul_Bengali (Lalon Fakir, Tagore Gitanjali, Bengali)

---

Notes on data quality and limitations

Script boundaries: The corpus contains texts in three scripts (Devanagari, Bengali, Latin/IAST). Character n-gram similarity across scripts is near-zero by design, not by historical finding. Cross-script comparisons should be interpreted with caution.

Contamination: Several large Bengali text files (Candidas, Bhakti Rasamrita Sindhu, Chaitanya Charitamrita) are archive.org full-text dumps including modern Bengali commentary alongside the original verses. The tradition labels reflect the source text, not necessarily the content purity of the file.

OCR quality: The Ramprasad Sen corpus (39,067 tokens) was OCR-processed from a 1914 scanned edition using Tesseract Bengali at 300 DPI. Estimated character error rate 5-10%, sufficient for term density analysis but not word-level morphological analysis.

Term lists: The Buddhist and Shakta vocabulary lists used for term density analysis are manually curated and reflect one set of choices about which terms belong to each tradition. Many terms have been used in both traditions historically. See the GitHub repository for the full term lists and methodology.

No significance testing: Cosine similarity scores are point estimates without confidence intervals. Bootstrap resampling for significance testing is planned for a future version.

---

Intended uses

• Diachronic Bengali language study: the corpus spans Old Bengali (8th c.) through modern Bengali (19th c.), enabling vocabulary evolution analysis across seven centuries.

• Computational study of religious syncretism in South Asia.

• Baseline vocabulary resource for Bengali historical NLP, particularly for the pre-modern period (8th-18th c.) which is underrepresented in existing Bengali NLP datasets.

• Control corpus for Sanskrit NLP: provides tradition-labeled Sanskrit texts across Buddhist, Shakta, and Vaishnava lineages.

• Training data for tradition classification models.

---

Citation

If you use this dataset, please cite:

@dataset{bose2026bengaldharma,
  author = {Bose, Joy},
  title = {Bengal Dharma Corpus: Evolution of Bengali Devotional Language
           across Buddhist, Shakta, and Vaishnava Traditions, 8th-19th c.},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/joyboseroy/bengal-dharma-corpus}
}

Related work:

@misc{bose2026darshana,
  author = {Bose, Joy},
  title = {Darshana-Graph: A Knowledge Graph of Indian Philosophical Traditions},
  year = {2026},
  eprint = {2606.18222},
  archivePrefix = {arXiv}
}

---

Sources

Sanskrit texts: sanskritdocuments.org, GRETIL (Gottingen Register of Electronic Texts in Indian Languages), Digital Sanskrit Buddhist Canon.

Bengali texts: archive.org (Ramprasad Sen 1914, Chaitanya Charitamrita, Bhakti Rasamrita Sindhu, Srikrishna Kirtan), Bengali Wikisource (Lalon Fakir), Society for Natural Language Technology Research (Gitanjali).

Gitagovinda: sanskritdocuments.org chapter-by-chapter Unicode edition.

Charyapada: Reconstructed from the standard scholarly editions (Haraprasad Shastri 1916, Per Kvaerne 1977).