A new AI-based deep learning technique has recovered ancient Greek texts, determined they date to the 5th century AD, and pinpointed their original location with an unprecedented precision.
According to Agence France Press (AFP), this technique described in the journal Nature, allows historians specializing in epigraphy to track tens of thousands of inscriptions engraved in stone, clay or metal.
Many of these inscriptions have deteriorated over time, leaving some text unreadable due to missing pieces or transfer from original site, and therefore, the radiocarbon dating technique cannot be used in this case.
To help epigraphists decipher these inscriptions, researchers from the Universities of Venice, Oxford, Athens in collaboration with Google’s DeepMind lab have developed a deep learning tool, an artificial intelligence technique that uses a “neural network” that simulates the human brain.
Named Ithaca, after the island of Odysseus in “The Iliad and The Odyssey”, this tool was trained on nearly 80,000 texts from the Packard Humanities Institute database, the largest digital collection of ancient Greek inscriptions. Ithaca’s language processing technique considers the order in which words appear in sentences and their links to each other to better contextualize them.
Because the texts feature many gaps, Ithaca had to merge the words and characters scattered on the stones. It then examined decrees from the 5th century BC engraved on stones from the Acropolis of Athens.
The tool assumed that the letter sequencing could help fill in the gaps in accordance with the historical context. For example, it suggested the word “covenant” to fill a six-character word missing from an oath of allegiance to a city in Athens. Then, the final decision to select the most credible prediction was left to the historians.
But their work was made much easier, as the work of Ithaca alone was 62% accurate. And when used by historians, the accuracy rate of the tool, described as“accessible”, jumped from 25% to 72%, explained the study published in the journal Nature, highlighting the benefits of man-machine cooperation.
