Deep Aramaic: Towards a synthetic data paradigm enabling machine learning in epigraphy.

Area of Science:

• Digital Humanities
• Computational Linguistics
• Artificial Intelligence in Archaeology

Background:

• Epigraphy faces challenges with limited labeled data for training machine learning (ML) algorithms, particularly for ancient scripts like Old Aramaic.
• Current ML techniques are constrained by the scarcity of real-world inscription data, hindering the analysis of damaged historical texts.

Purpose of the Study:

• To pioneer a novel methodology for generating synthetic training data specifically for Old Aramaic letters.
• To overcome the limitations imposed by scarce labeled data in epigraphic analysis.
• To enhance the accuracy of interpreting damaged ancient inscriptions.

Main Methods:

• Developed a pipeline to synthesize photo-realistic Old Aramaic letter datasets, incorporating diverse textural features, lighting conditions, damage, and augmentations.
• Engineered a large corpus of 250,000 training and 25,000 validation images covering all 22 Aramaic letters.
• Trained a residual neural network (ResNet) model on the synthetic dataset for classifying degraded Aramaic letters.

Main Results:

• The ResNet model achieved 95% accuracy in classifying real Old Aramaic letters from an 8th century BCE inscription.
• Validated the model's effective generalization across varying materials and styles.
• Demonstrated the model's capability to handle diverse real-world scenarios with degraded inscriptions.

Conclusions:

• The synthetic data generation approach is viable and effectively overcomes the dependence on scarce training data in epigraphy.
• The innovative framework significantly enhances interpretation accuracy for damaged inscriptions.
• This methodology advances knowledge extraction from historical epigraphic resources.