Sanskrit Dating

Category-free relative chronology of Sanskrit texts, inferred from intertextual citation structure together with stylometric signal (morphology + function-word n-grams), and resolved into absolute dates with a Gibbs sampler anchored on a set of externally-dated works.

🗺️ Interactive chronology map

An interactive, searchable timeline of the dated corpus is published via GitHub Pages:

→ https://dharmamitra.github.io/sanskrit-dating/

The map (sanskrit_chronology_interactive.html) lets you hover for per-work dates and notes, search by title / author / anchor text, and zoom into any period. Static renderings are also included (sanskrit_chronology.png, sanskrit_chronology_all.png).

Approach

1. Intertextual graph — parallel passages between works (matches/) are extracted into a directed citation/relation graph (extract_edges.py, extract_relations.py, orient.py, mst.py). Direction of borrowing is the key signal; an undirected graph alone does not date.
2. Stylometric signal — morphology and function-word n-grams give a genre-largely- independent "composition-style" clock (extract_morph.py, extract_ngrams.py, extract_pos.py, date_morph*.py). Raw vocabulary turned out to be too weak a clock on its own (Sanskrit is normatively frozen).
3. Anchoring & inference — known/argued dates are encoded as priors and order constraints (manual_constraints.tsv, researched_anchors.tsv, chronbmm_priors.tsv, vedic_anchors.tsv, dcs_anchors.tsv). A Gibbs sampler propagates dates over the graph subject to these constraints (date_gibbs*.py, date_gibbs_full.py).
4. Visualization — visualize*.py render the results to PNG and the interactive HTML.

Repository layout

Path	What it is
extract_*.py	Feature/edge/relation extraction from the corpus
dating*.py, date_*.py	Dating models (morphology, directional graph, Gibbs sampler)
visualize*.py	Render PNG + interactive HTML chronology
*.tsv	Constraints, anchors, and dated results (e.g. dated_gibbs_full.tsv)
meta.json, metadata.tsv, text-information.json, text_info/	Work metadata
sanskrit_chronology_interactive.html	The interactive map (served by Pages)
*.png	Static chronology renderings

Key result files

• dated_gibbs_full.tsv — final dated works from the full Gibbs run (input to the map)
• dated_chunks_works.tsv — per-work date estimates from chunk-level dating
• researched_anchors.tsv, manual_constraints.tsv — the anchoring evidence base
• sanskrit-editions.json — for each text, its printed/critical edition (traced from the GRETIL/DSBC source header) and a link to the scan on archive.org where one was found (1494/1618 editions identified; 833 with a confident archive.org match)

Data not included

Large, regenerable artifacts are not version-controlled (see .gitignore):

• matches/ — the raw parallel-passage corpus (~2.2 GB)
• *.pkl — extracted feature stores (vocab.pkl, chunks_fw.pkl, fwgrams.pkl, …)
• chunks_dense.tsv — dense chunk matrix (~62 MB)

These are produced by the extract_*.py scripts from the source corpus and feature stores.

Reproducing

# 1. extract features + intertextual edges from the corpus (regenerates the .pkl / matrix files)
python extract_chunks.py
python extract_morph.py
python extract_edges.py
python extract_relations.py

# 2. run the dating model
python date_gibbs_full.py        # -> dated_gibbs_full.tsv

# 3. build the visualizations
python visualize.py              # -> sanskrit_chronology.png
python visualize_interactive.py  # -> sanskrit_chronology_interactive.html