Convergent Scripture, Divergent Rulings

Text-mining the muʿāmalāt verses across five Southeast Asian Qur'an editions.

This repository reproduces the analysis behind the paper Convergent Scripture, Divergent Rulings: How Southeast Asian States Translate and Rule on the Economic Qur'an (submitted to Studia Islamika). It transposes the text-mining method of Lange, van Kuppevelt & van der Zwaan (2021), Text Mining Islamic Law, from a mono-lingual, diachronic Arabic corpus to a multilingual, synchronic corpus of national Qur'an translations.

Core finding — the two layers decouple. Translation is a layer of convergence: cross-edition similarity is governed by language family, and the Malay editions of Malaysia and Brunei are the most similar pair (mean cosine 0.92). Fatwa is the layer of divergence: on bank interest, insurance, cryptocurrency, and tawarruq the same states run from most to least permissive as Malaysia, Singapore, Brunei, Indonesia — an ordering their translations do not predict.

Method: Lange et al. → this study

Lange et al. (2021)	This study
Text-reuse network (detects shared citations)	Cross-edition semantic similarity (LaBSE cosine) — reuse is trivial when all editions render one Arabic source
word2vec trained on the corpus (mono-lingual)	LaBSE language-agnostic sentence embeddings (5 languages in one space)
Parsimonious word clouds (weighwords)	Parsimonious language model (Hiemstra et al. 2004), same instrument, on the Malay-family trio
Frequency analysis (BlackLab)	Frequency word cloud of the muʿāmalāt corpus
LDA topic modeling	Per-domain divergence (105 verses is too sparse for stable LDA)

Repository layout

data/
  economic_verses_aligned.csv     # the corpus: 105 verses x 5 editions (+ Arabic, footnotes)
  fatwa_comparison.csv            # fatwa strictness coding, 4 issues x 4 states
  out/                            # derived data (committed, so the notebook runs offline)
    embeddings.npy                #   LaBSE vectors, 525 x 768  ([id,ms,th,tl,bn] x 105)
    lang_similarity_matrix.csv    #   5x5 mean-cosine matrix
    domain_divergence.csv, pairwise_long.csv, verse_divergence.csv, ...
    figs/                         #   all figures (PNG)
    tables/                       #   publication tables (CSV/HTML/LaTeX)
scripts/                          # the pipeline (see below)
notebooks/
  reproduce_analysis.ipynb        # self-contained showcase: matrix, divergence, stats, two-layer
manuscript/                       # the paper (gitignored by default)
requirements.txt                  # light stack (reproduce from cached data)
requirements-embeddings.txt       # heavy stack (recompute LaBSE from scratch)
install.R                         # R packages for figures/tables
run_all.sh                        # regenerate everything end to end

Quick start (5 minutes, no GPU, no model download)

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook notebooks/reproduce_analysis.ipynb

The notebook loads the cached embeddings.npy, recomputes the similarity matrix (it asserts it matches the reported values), reproduces the domain-divergence chart, re-runs the Wilcoxon and Kruskal–Wallis tests in Python, and plots the two-layer decoupling — all from committed data, no network access required.

Full pipeline

The pipeline is a Python → CSV → R hand-off. Run it end to end with:

bash run_all.sh

or step by step:

Step	Script	Output
1	01_fetch_corpus.py	rebuilds economic_verses_aligned.csv (needs network)
2	03_embed_sim.py	LaBSE embeddings + similarity/divergence CSVs (needs heavy stack)
3	04_visualize.R	fig1–4, publication tables
4	05_stats.R	bootstrap CI + Wilcoxon + Kruskal (fig5)
5	06_network.R	fig6 verse similarity network
6	12_parsimonious_wordcloud.py → 07_wordcloud.R	fig7 (parsimonious) + fig8 (thematic)
7	08_fatwa_heatmap.R	fig9 fatwa-strictness heatmap
8	10_python_figures.py, 11_more_figures.py	supplementary figures (t-SNE, choropleth, …)
9	09_build_docx.py	assembles the manuscript (optional)

Data

See DATA_SOURCES.md for the provenance of each edition and an important caveat: Indonesia/Malaysia/Brunei are state-authorised editions, whereas Thailand and the Philippines use the open Rowwad editions (no clean state text is available). Translation strings remain the property of their authorities and are included only as verse-level scholarly excerpts.

Citation

If you use this code or corpus, please cite the paper (see CITATION.cff) and the method it builds on:

Lange, Christian, Dafne van Kuppevelt, and Janneke van der Zwaan. 2021. "Text Mining Islamic Law." Islamic Law and Society 28: 234–281.

License

Code: MIT (LICENSE). Translation data: property of the respective authorities, verse-level excerpts only. See the note in LICENSE and DATA_SOURCES.md.