Contextualizing ancient texts with generative neural networks

Area of Science:

• Digital Humanities
• Artificial Intelligence in History

Background:

• Ancient inscriptions offer direct insights into historical civilizations.
• Current digital methods for analyzing inscriptions are limited to literal matches and narrow scopes.
• Historians rely on identifying textual parallels for contextualization, restoration, and attribution.

Purpose of the Study:

• To introduce Aeneas, a generative neural network designed for contextualizing ancient texts.
• To leverage AI for retrieving textual and contextual parallels, utilizing visual inputs, and restoring texts.
• To advance the state-of-the-art in historical text analysis tasks.

Main Methods:

• Development of Aeneas, a generative neural network for ancient text contextualization.
• Integration of textual and visual data processing capabilities.
• Evaluation through a large-scale study involving historians using Aeneas outputs.

Main Results:

• Historians found Aeneas-retrieved parallels useful in 90% of cases, boosting confidence by 44%.
• Aeneas-assisted historians outperformed humans and AI alone in restoration and geographical attribution.
• Aeneas achieved a 13-year accuracy in dating inscriptions compared to ground-truth ranges.

Conclusions:

• Aeneas significantly enhances historical research workflows by providing contextual parallels and aiding in text analysis.
• The integration of AI with historical methods offers transformative tools for understanding the past.
• Aeneas demonstrates the potential of AI to advance the study of ancient civilizations through inscription analysis.