DeepMind AI Beats Human Historians at Deciphering Ancient Texts

University of Oxford researchers are using machine learning to help make ancient text restoration less tedious.
Ancient Greek tablet in need of restoration
DEA / L. PEDICINI / Contributor

It’s pretty easy to figure out what I mean if __ few symbols are missing from this sentence. But what __ ____ are _______ missing? Context might allow you to guess the content of the previous sentences, but could you do the same if ___ _____ ___ _______ _______?

According to a new paper by researchers from DeepMind and the University of Oxford’s Faculty of Classics, AI can help restore, understand, and recreate ancient Greek texts that have been damaged and left with gaps that make them nearly impossible to understand. The work will be presented next month at the Empirical Methods in Natural Language Processing conference in Hong Kong.


The researchers used an algorithm named after Pythia, the woman who in Greek mythology was a vessel for Apollo’s prophecies. They found that it outperformed historians trained in restoring fragmented stone, clay, or metal tablets. While the historians hit about 43% accuracy after two hours, Pythia landed a nearly 70% accuracy rate after needing a few seconds.

It’s often difficult for historians to accurately restore ancient tablets because entire sections of symbols may have been wiped away over the years. Epigraphy, the field of study dedicated to piecing together the content of inscriptions using physical and environmental context, has limitations such as a historian’s own biases and fabrications. These factors, plus the complexity of accurately dating inscriptional evidence, all make epigraphic work incredibly difficult and time-consuming.

Importantly, Pythia, was created not to replace historians doing the work but to assist them. The algorithm was trained to look at a damaged text, predict its character sequences, then provide those predictions as suggestions with an accompanying confidence level.

To train Pythia, the researchers converted the world's largest digital collection of ancient Greek inscriptions into a format that their neural network could understand. The training data also included information on things like the shape of fragmented symbols, the preceding and upcoming syntax and semantics of symbols or sections, and the layout of similar texts. The ability to use contextual clues to make predictions was what really allows Pythia to assist in epigraphical work.

The rest falls to the historian, who can use their knowledge and understanding to make the best pick from among the options offered by Pythia.

While Pythia could assist in radically improving ancient text restoration, it's just that—assistance. Humans still need to perform the archaeological excavation, piece together the physical inscriptions as best as they can, make out what the symbols, and decide which of Pythia’s guesses make the most sense.