Armenian Epigraphy NLP Pipeline

This repository contains a research-oriented NLP pipeline for processing Armenian epigraphic corpora, extracting structured inscription data, and detecting inscription-type terminology using Named Entity Recognition (NER).

The project focuses on OCR-heavy, layout-complex historical sources and low-resource Armenian language settings.

Repository Structure

• Generated_data/
  Model outputs, predictions, and GT-labeled results

• Inference/
  Contains Inference.ipynb for running trained models on unseen data

• Notebooks/
  Experimental notebooks for training and evaluation. This is an aggregation of all the work that has been done. Main Folder to look at.

• dataset_formation/
  Dataset construction, IOB tagging, and train/dev/test splits

• preprocessing/
  OCR cleaning, paragraph restructuring, and sentence splitting

• images/
  Figures used for analysis and reporting

• corpus_inscription_terms.xlsx
  Expert-curated list of Armenian inscription-type terminology

• progressive_final.pt Pt file for finetuned NER model

Main Features

• OCR-aware text cleaning and restructuring
• Rule-based inscription extraction with metadata
• GPT-assisted ground-truth generation (JSON via Pydantic)
• NER training with:
  • spaCy baseline
  • Transformer models (mBERT, XLM-RoBERTa, Armenian-specific)
• Evaluation on seen vs unseen terminology

Typical Workflow

1. Preprocess OCR text (preprocessing/)
2. Form NER-ready datasets (dataset_formation/)
3. Train and evaluate models (Notebooks/)
4. Run inference on new corpora (Inference/)
5. Inspect outputs (Generated_data/)

Scope & Limitations

• Strong generalization to known inscription types in new contexts
• Limited discovery of entirely unseen terminology
• Designed as a research prototype, not a production system

Use Cases

• Digital epigraphy
• Cultural heritage digitization
• Low-resource NLP research
• Historical corpus structuring

License

See LICENSE for details.