Skip to content

Unsupervised NLP Topic Modeling for Classical Arabic

Case Study Summary

Type: Independent project
Domain: Natural Language Processing
Corpus: 6,236 verses of classical Arabic

Highlights:

  • Transformer sentence embeddings built for Arabic: GATE-AraBERT-v1 and Arabic LaBSE-Matryoshka
  • UMAP dimensionality reduction and HDBSCAN density-based clustering
  • BERTopic-style topic modeling to name and describe each theme
  • Automated PDF reporting that explains the findings to a general audience

Classical Arabic is a hard target for NLP. Its vocabulary, morphology and style differ from the modern text most language models are trained on. I wanted to find out whether an unsupervised pipeline could surface meaningful themes across a full classical corpus, without any labeled data.

Challenge

  • No labels. The structure had to emerge from the text itself.
  • Language fit. Generic multilingual embeddings often miss the nuance of classical Arabic.
  • Explainability. The output had to make sense to readers who aren't data scientists.

Approach

I built a multi-phase Python pipeline:

  1. Embedding. Encode every verse with transformer sentence embeddings tuned for Arabic (GATE-AraBERT-v1 and Arabic LaBSE-Matryoshka) and compare how well each captures meaning.
  2. Dimensionality reduction. Project the high-dimensional embeddings with UMAP while preserving local structure.
  3. Clustering. Use HDBSCAN to find clusters of varying density, letting outliers stay outliers instead of forcing them into a group.
  4. Topic modeling. Apply BERTopic-style class-based term weighting to describe each cluster as an interpretable topic.
  5. Reporting. Generate PDF reports automatically, so the results can be read by a general audience.

Solution Overview

flowchart TD
    A[6,236-verse<br/>Arabic corpus] --> B[Sentence embeddings<br/>GATE-AraBERT / LaBSE]
    B --> C[UMAP<br/>reduction]
    C --> D[HDBSCAN<br/>clustering]
    D --> E[BERTopic-style<br/>topic modeling]
    E --> F[Automated<br/>PDF report]

What this demonstrates

  • Choosing and evaluating embedding models for a specific, low-resource language domain
  • Unsupervised learning end-to-end: representation, reduction, clustering and interpretation
  • Turning technical NLP results into material a non-technical audience can read

Tech Stack

  • Python
  • Sentence Transformers (GATE-AraBERT-v1, Arabic LaBSE-Matryoshka)
  • UMAP
  • HDBSCAN
  • BERTopic-style topic modeling
  • Automated PDF report generation

  • Let's have a virtual coffee together!

    Want to see if we're a match? Let's have a chat and find out. Schedule a free 30-minute session to discuss your data and AI challenges and explore how we can work together.

    Book Free Intro Call