The Patrologia Graeca Corpus: OCR, Annotation, and Open Release of Noisy Nineteenth-Century Polytonic Greek Editions

This paper introduces the Patrologia Graeca Corpus, a large-scale open resource featuring OCR-processed, lemmatized, and part-of-speech tagged text from degraded nineteenth-century bilingual Greek-Latin editions, which achieves state-of-the-art recognition accuracy and establishes a new benchmark for noisy polytonic Greek processing.

The Gist

Imagine a massive, ancient library called the Patrologia Graeca. It's like a time capsule containing 161 giant volumes of Greek and Latin writings from the early days of Christianity up to the Middle Ages. For over a century, these books have been sitting in a digital "dark room." They exist as scanned PDFs (basically pictures of pages), but you can't search them, copy the text, or ask a computer to analyze them. It's like having a library where you can see the books on the shelf, but the pages are glued shut.

The problem? These books were printed in the 1800s with messy, fancy Greek letters (called "polytonic" Greek) that have tiny accents and marks. Over time, the paper got old, the ink faded, and the scans got blurry. Trying to use a standard computer program to read them is like trying to read a handwritten note written in a storm while wearing foggy glasses. The computer gets confused, mixes up letters, and produces garbage text.

Enter the "Patrologia Graeca Corpus" project.

Think of the authors of this paper as a team of digital archaeologists and master librarians. They didn't just want to take a photo of the books; they wanted to unlock the words inside so anyone could use them. Here is how they did it, using some creative metaphors:

1. The "Smart Eye" (Layout Detection)

First, they had to teach the computer how to see the page structure. Imagine a page from these old books is a busy city street. There are two main lanes: one for Greek and one for Latin, often weaving around each other. There are also side streets (marginal notes) and big billboards (titles).

Old computers got lost in this traffic. The team built a YOLO-based detector (think of it as a super-smart security camera). This camera doesn't just look at the whole street; it instantly spots exactly where the Greek lane starts and ends, ignoring the Latin lane and the messy notes in the margins. It draws a digital "fence" around the Greek text so the computer knows exactly what to read.

2. The "Translator with Training Wheels" (OCR)

Once the computer knows where to look, it has to read the letters. This is the hardest part. The Greek letters are like a set of twins that look almost identical, except for a tiny accent mark on their head. A standard computer might think a "smooth breathing" mark is a "rough breathing" mark, changing the meaning of the word entirely.

The team trained a CRNN model (a type of AI that reads text). To teach it, they didn't just show it clean, perfect books. They took clear text and intentionally ruined it with digital "noise"—adding fake scratches, blurring, and fog to mimic the old, damaged books. They taught the AI to recognize the letters even when they were dirty or blurry.

The Result?
Before this project, the best computers got about 90% of the words right (which is actually quite bad for research). This new system got 99% of the characters right and 95% of the words right. It's the difference between a student who guesses the spelling of a word and a scholar who knows the exact definition.

3. The "Dictionary Detective" (Linguistic Analysis)

Getting the text right is only step one. Ancient Greek is a language where words change shape depending on how they are used (like "go," "went," "gone"). The team didn't just stop at reading the words; they added a linguistic layer.

Imagine taking every word in the book and attaching a digital tag that says:

• "This is the word 'love'."
• "This is the past tense."
• "This is the subject of the sentence."

They did this for 6 million words. They created a "clean" version of the text where you can search for "love" without worrying about whether the computer is looking for "loved" or "loves."

Why Does This Matter?

Think of this project as turning a locked vault into a public park.

• For Historians: They can now search for specific rare words or themes across thousands of pages instantly, something that used to take years of manual reading.
• For AI: This is a goldmine of training data. Just as you need to read many books to learn a language, AI models need massive amounts of text to learn. This project gives future AI models a huge, high-quality "textbook" of Ancient Greek, helping them understand the language better than ever before.
• For Everyone: The data is free and open. Anyone with an internet connection can go to their website, search the texts, and explore the history of the ancient world without needing a PhD or a library card.

In a nutshell: The team took a messy, unreadable pile of 19th-century Greek books, taught a computer how to see through the noise, cleaned up the text, added smart tags, and opened the doors to the public. They turned a "dark room" library into a bright, searchable, and living digital treasure.

Technical Summary

Here is a detailed technical summary of the paper "The Patrologia Graeca Corpus: OCR, Annotation, and Open Release of Noisy Nineteenth-Century Polytonic Greek Editions."

1. Problem Statement

The Patrologia Graeca (PG), a massive 161-volume collection of patristic and Byzantine Greek texts compiled by Jacques-Paul Migne (1857–1866), remains largely inaccessible for computational linguistics.

• Current State: Most volumes exist only as non-searchable PDF scans with complex, heterogeneous layouts (bilingual Greek-Latin columns, marginalia, running titles).
• Limitations of Previous Efforts:
  • Existing digital initiatives (e.g., Open Patrologia Graeca 1.0) produced OCR outputs with high error rates (typically <90% accuracy), inconsistent encoding, and no linguistic annotation.
  • Standard OCR tools (Tesseract, Transkribus) fail on PG due to degraded typography, complex polytonic diacritics, bilingual interference, and scanning artifacts (curvature, misalignment).
  • There is a lack of structured, open, and linguistically annotated corpora for post-classical and Byzantine Greek, hindering the training of modern Large Language Models (LLMs) and NLP tools for this period.

2. Methodology

The authors developed a dedicated pipeline combining advanced layout detection, iterative OCR fine-tuning, and hybrid linguistic annotation.

A. Data Preparation & Ground Truth

• Test Set: A manually transcribed 30-page test set was created to evaluate performance.
• Training Data: 445 images representing the typographical diversity of the PG were semi-automatically aligned using the Calfa Vision platform. This included manual correction of automatic layout detections, with a specific focus on pages with crossing text regions (Greek/Latin overlap).
• Augmentation: The foundation model was trained on synthetic data augmented with aggressive degradations (Gaussian noise, motion blur, compression artifacts, elastic distortion) to simulate the visual variability of 19th-century scans.

B. OCR Pipeline

• Layout Detection: Utilized YOLO-based models to identify semantic zones (Greek columns, Latin columns, titles, marginalia) and reading order. This visual approach outperformed textual/hybrid models for complex layouts.
• Text Recognition: Employed a CRNN (Convolutional Recurrent Neural Network) architecture.
  • Pre-training: Started with a model trained on the Genavensis Græcus 44 manuscript.
  • Fine-tuning: Iteratively fine-tuned on the PG-specific dataset using an active learning strategy (correcting errors and re-injecting them into training).
• Linguistic Analysis:
  • Lemmatization & POS Tagging: Used a hybrid strategy combining neural tagging (based on the PIE architecture adapted for the GREgORI tagset) with rule-based/dictionary post-correction.
  • Normalization: Implemented Unicode normalization to handle diacritic inconsistencies (e.g., monotonic vs. polytonic forms).

C. Corpus Construction

• Filtering: Automated removal of Latin text, hyphens, and empty lines.
• Output Format: Texts were structured in the .vert format (used by Sketch Engine), containing five layers of annotation per token: OCR wordform, intuitive form (lowercase, diacritic-free), lemma, intuitive lemma, and morpho-syntactic tag.
• Traceability: The output preserves document, page, and line IDs to link OCR results back to the original PDF context.

3. Key Contributions

1. First Large-Scale Open Resource: Release of the first open, linguistically enriched corpus for the undigitized volumes of the Patrologia Graeca (~6 million lemmatized tokens).
2. State-of-the-Art OCR Performance: Achievement of a Character Error Rate (CER) of 1.05% and a Word Error Rate (WER) of 4.69% on noisy, polytonic Greek, significantly outperforming existing systems.
3. Benchmark Dataset: Creation of a high-quality ground truth dataset (445 images, 11,096 text lines) specifically for 19th-century polytonic Greek, addressing the lack of training data for this domain.
4. Hybrid Annotation Workflow: Demonstration of a robust pipeline combining neural models with rule-based correction for handling highly inflected and diachronic Greek.
5. Public Accessibility: Full release of raw OCR data, layout annotations, and structured .vert files via GitHub and Zenodo, with a searchable interface on gregoriproject.com.

4. Results

• OCR Accuracy:
  • Ours (PG fine-tuned): CER 1.05%, WER 4.69%.
  • Transkribus (19th c. Greek): CER 6.14%, WER 14.82%.
  • Tesseract (Greek): CER 11.57%, WER 39.65%.
  • Improvement: A gain of 5–7 percentage points in CER and 6–10 in WER over previous baselines.
• Error Analysis:
  • Primary Errors: Diacritic confusions (e.g., distinguishing between monotonic and polytonic accents on the same letter) account for >80% of errors.
  • Secondary Errors: Spacing, punctuation, and confusion between visually similar characters (e.g., $\iota$ vs. $\tau$ in poor ink regions).
  • Layout: YOLO achieved high precision (mAP50 > 0.97) for main text columns but struggled slightly with titles and marginalia due to script ambiguity.
• Corpus Characteristics:
  • Lexical Diversity: The corpus introduces thousands of rare inflected forms, technical/theological terms, and named entities underrepresented in existing corpora.
  • Visual Distinctiveness: t-SNE visualizations confirm the PG corpus occupies a unique typographic and lexical space, distinct from manuscripts and modern prints.

5. Significance

• Philological Impact: Makes a vast portion of Byzantine literature machine-readable and searchable for the first time, enabling large-scale quantitative analysis of texts that have not been re-edited since the 19th century.
• NLP & AI Advancement: Provides essential training material for next-generation Ancient Greek language models (e.g., Ancient-Greek-BERT, GreBERTa), improving their ability to handle diachronic variation and complex morphology.
• Methodological Blueprint: Establishes a reproducible workflow for digitizing "noisy" historical documents with complex layouts and polytonic scripts, serving as a model for similar projects in digital humanities.
• Open Science: By releasing the data and models openly, the project democratizes access to high-quality Greek philological resources, fostering further research in historical OCR and cross-lingual analysis.