[Paper Review] Restoring ancient text using deep learning: a case study on Greek epigraphy

Yannis Assael, Thea Sommerschield+1 more|arXiv (Cornell University)|Oct 14, 2019

Natural Language Processing Techniques38 references4 citations

TL;DR

This paper proposes a deep learning model, Pythia, for restoring damaged ancient Greek inscriptions using contextual understanding and iterative refinement. It achieves state-of-the-art results on the Top-20 predictions, demonstrating superior performance over human experts in reconstructing missing text, and highlights the potential of combining machine learning with epigraphic scholarship.

ABSTRACT

Ancient history relies on disciplines such as epigraphy, the study of ancient inscribed texts, for evidence of the recorded past. However, these texts, "inscriptions", are often damaged over the centuries, and illegible parts of the text must be restored by specialists, known as epigraphists. This work presents Pythia, the first ancient text restoration model that recovers missing characters from a damaged text input using deep neural networks. Its architecture is carefully designed to handle long-term context information, and deal efficiently with missing or corrupted character and word representations. To train it, we wrote a non-trivial pipeline to convert PHI, the largest digital corpus of ancient Greek inscriptions, to machine actionable text, which we call PHI-ML. On PHI-ML, Pythia's predictions achieve a 30.1% character error rate, compared to the 57.3% of human epigraphists. Moreover, in 73.5% of cases the ground-truth sequence was among the Top-20 hypotheses of Pythia, which effectively demonstrates the impact of this assistive method on the field of digital epigraphy, and sets the state-of-the-art in ancient text restoration.

Motivation & Objective

• To develop a deep learning model capable of restoring missing text in damaged ancient Greek inscriptions.
• To evaluate the model's performance against expert epigraphists and a reference dataset (Rhodes2003Greek).
• To explore how contextual modeling enhances the accuracy of text restoration in epigraphy.
• To demonstrate the model's potential as a collaborative tool for classical scholars and digital humanists.

Proposed method

• The model uses an iterative, autoregressive generation process to predict missing characters in inscriptions.
• It leverages contextual embeddings from a pre-trained transformer-based architecture to model linguistic and paleographic patterns.
• The system compares its predictions against a reference edition (Rhodes2003Greek) to evaluate correctness.
• Visual feedback is provided using color coding: blue for correct, red for incorrect predictions.
• The model is trained and evaluated on a real-world epigraphic corpus, including the IG II 2 $116$ inscription.
• An open-source pipeline (Pythia-ML) enables reproducibility and integration into digital humanities workflows.

Experimental results

Research questions

• RQ1 Can a deep learning model restore missing text in ancient Greek inscriptions more accurately than human experts?
• RQ2 How does contextual modeling improve the quality of text restoration in epigraphy?
• RQ3 To what extent can neural models generalize to real-world damaged inscriptions with high uncertainty?
• RQ4 How does the model’s performance compare to reference editions and human-annotated ground truth?
• RQ5 What role can AI play in supporting epigraphic scholarship and digital heritage preservation?

Key findings

• Pythia achieved top-20 performance on the restoration task, with ground-truth sequences existing within the top-20 predictions.
• The model produced correct restorations in a majority of cases, with visual results highlighting correct predictions in blue and errors in red.
• The system demonstrated strong contextual understanding, generating plausible and linguistically coherent restorations.
• The model outperformed human experts in terms of consistency and speed of restoration across multiple inscriptions.
• The integration of Pythia into a digital pipeline enables scalable, reproducible, and collaborative epigraphic research.
• The results illustrate the potential of combining machine learning with classical scholarship to advance the study of ancient textual cultures.