AI can help historians restore ancient texts from damaged inscriptions

An AI tool developed by DeepMind can help historians restore ancient Greek texts with 72 per cent accuracy, and date inscriptions to within 30 years of their true age

By Carissa Wong

An artificial intelligence algorithm developed as part of a collaboration between historians and UK-based AI firm DeepMind can help restore ancient Greek texts with 72 per cent accuracy.

The AI can also predict where in the ancient Mediterranean world the texts were originally written with more than 70 per cent accuracy and date them to within a few decades of their agreed-upon date of creation. All of this marks an improvement upon an earlier version of the AI that could only restore ancient texts.

“Inscriptions provide evidence of the thought, language, society and history of past civilisations,” says Thea Sommerschield at Ca’ Foscari University of Venice in Italy. “But most surviving inscriptions have been damaged over the centuries, so their texts are now fragmentary or illegible. They may also have been moved or trafficked far from their original location.”

When recovering ancient texts, historians are usually interested in achieving three major goals: restoring the text, and working out exactly when and where it was written. To do this, they look for distinctive features and patterns in the style of writing and compare them to those of ancient texts that have already been found and dated.

“However, it’s really difficult for a human to harness all existing relevant data, and to discover underlying patterns every time,” says Sommerschield.

Sommerschield and her colleagues worked with researchers at DeepMind to get the machine-learning AI – called Ithaca after a Greek island that is famous for being the home of the legendary figure Odysseus – to carry out all three tasks.

To train Ithaca, the team used around 60,000 ancient Greek texts from across the Mediterranean that are already well-studied and known to have been written between 700 BC and AD 500. The team masked some of the characters in the texts and then compared Ithaca’s predictions for this “missing” text with the actual inscriptions.

Next, the team used a data set of nearly 8000 inscriptions – again, already well-studied and understood – to test Ithaca’s performance alone, or in combination with two ancient historians. On its own, Ithaca could restore texts with 62 per cent accuracy, while ancient historians alone restored text with around 25 per cent accuracy.

However, the most accurate reconstructions involved Ithaca and historians working together. When historians took Ithaca’s top 20 most likely reconstructions for a given text and used them to inform their own work, they could restore the text with an accuracy even greater than Ithaca alone.

“When historians used Ithaca, their performance on the text restoration task actually tripled, to 72 per cent,” says Sommerschield.

Ithaca could also predict where in the Mediterranean a text was written 71 per cent of the time and it could date the texts to within 30 years of their true date of creation, as previously established by historians.

“It is clear that the authors’ work is important and groundbreaking. The ‘ancient historian and Ithaca’ method produces startlingly significant improvements in outcomes over traditional human-only methods,” says Tom Elliott at New York University. However, further testing with more historians is needed and people will need training and technical support to use the tool, he adds.

The team says the feedback from historians so far has been positive.

“We hope that the way we’ve designed it, it’s going to be easy for an ancient historian to use, because they will just type in the text [to an online interface] and then they will get all these visualisations that they can use,” says Yannis Assael at DeepMind in the UK, and an author of the study.

Ithaca’s design should also make it easily applicable to any ancient language and any written medium, says Sommerschield.

Journal reference: Nature, DOI: https://doi.org/10.1038/s41586-022-04448-z