EccLexica : Christian Architectural Discourse Detection in Latin Texts

This repository contains a dictionary-based method for identifying architectural discourse in Late Antique and Medieval Latin texts. It is applied to a test corpus of Latin inscriptions.

The workflow combines a dictionary-based approach with an evaluation against GLiNER named-entity recognition models.

EccLexica is being developed as part of the ANR E-cclesia project.

---

Repository Structure

.
├── data/
│   └── dicts
│     └── auto_terms.csv
│     └── asso_terms.csv
│     └── mat_terms.csv
│   └── sample.csv
│
├── output/
│
├── arch-score.py
├── gliner-eval.py
└── README.md

---

Data

• Main dataset

The main dataset is not stored in this repository due to its size.

Please download EDCS_text_cleaned_2022-09-12.json from Zenodo:

👉 https://zenodo.org/records/7072337

After downloading, place the file in the following location:

data/EDCS_text_cleaned_2022-09-12.json

• Sample dataset (data/sample.csv)
  • 100 inscriptions.
  • Used for testing and GLiNER evaluation.

---

Scripts

arch-score.py

Computes an architectural concentration score (0–100) for each inscription based on the presence and distribution of architectural vocabulary.

Main steps:

1. Filtering
   • Date range: 399–1199 CE.
   • Excludes a predefined list of eastern provinces.
2. Lemmatization
   • Uses spaCy’s Latin model (la_core_web_md) on cleaned interpretive text.
3. Scoring
   • Uses three term categories:
     • Autonomous (e.g. basilica, ecclesia)
     • Associative (e.g. murus, porta)
     • Material (e.g. marmor, aurum)
   • Score combines:
     • Term count
     • Co-occurrence of term types
     • Proximity of terms in the text
     • Term density
4. Output generation
   • Full scored dataset.
   • Score-based subsets (per 10-point bin).
   • High-score subset (score > 50).

---

gliner-eval.py

Evaluates GLiNER NER models against the dictionary-based architectural terms.

What it does:

• Loads a scored sample dataset from output/sample.csv.
• Runs multiple GLiNER models on:
  • Lemmatized text
  • Cleaned interpretive text
• Compares GLiNER-extracted entities with dictionary-based terms.

Evaluated labels:

• building/type
• building/part
• building/material

Metrics computed:

• Precision
• Recall
• F1 score
• Percentage and count of overlapping terms
• Number of inscriptions with at least one shared term

---

Requirements

• Python 3.9+
• Core libraries:
  • pandas, numpy, scikit-learn
  • spacy (+ la_core_web_md)
  • gliner

---

Typical Workflow

1. Run arch-score.py to filter inscriptions and compute architectural scores.
2. Inspect or subset results in the output/ directory.
3. Run gliner-eval.py on a scored sample to compare dictionary-based detection with GLiNER models.

---

Notes

• Scores are heuristic and intended for comparative and exploratory analysis, not absolute classification.

---

Citations

Heřmánková, Petra. “EDCS”. Zenodo, September 12, 2022. https://doi.org/10.5281/zenodo.7072337.