Artificial intelligence restores historical manuscripts

Researchers are developing a tool for deep learning, an artificial intelligence technique that uses a "neural network" that mimics the structure of the human brain.

Techniques to protect people's cultures

PARIS - A deep learning technology based on artificial intelligence has made it possible to recover, date and locate ancient Greek texts from the fifth century BC with unparalleled accuracy, according to the results of a study published in the journal Nature. Tens of thousands of inscriptions on stone, clay or metal allow inscription historians to trace the history of ancient civilizations.

But much of it has deteriorated over time, to the point that some texts are rendered unreadable with parts missing. These precious sources have also been moved far from their place of origin, which complicates their dating, as carbon-14 dating cannot be used on inorganic materials.

To help inscriptionists decipher them, researchers from the universities of Venice, Oxford and Athens and Google's DeepMind company have developed a tool for deep learning, an artificial intelligence technology that uses a "neural network" that mimics the structure of the human brain.

The Ithaca technique alone achieved an accuracy of 62 percent by incorporating individual words and letters dispersed throughout the texts

This tool, called "Ithaca," referring to Odysseus' island in "The Iliad and the Odyssey," was trained on nearly 80,000 vignettes referenced in the Packard Institute for the Humanities' database, the largest collection of ancient Greek numerical inscriptions.

Automatic language processing takes into account the order in which words appear in sentences and their relationships, in order to better contextualize them.

In light of the multiple gaps in the texts, Ithaca had to combine both the individual words and letters dispersed across the boards and the pieces on which it was inscribed. Then the tool dealt with a group of documents from the fifth century BC inscribed on stone tablets and found in the landmark of the Acropolis in Athens.

The tool assumed letter sequences that could fill in the gaps, consistent with the historical context. For example, the tool suggested the word "covenant" to fill in a gap of six missing letters in an oath of allegiance to a city in Athens.

Then it is up to historians to choose the most reliable prediction. Their work was greatly facilitated, as the Ithaca tool, presented as "available", alone achieved an accuracy of 62 percent. And when used by historians, this accuracy rate jumps from 25 to 72 percent, according to the study, whose results were published in the journal Nature, highlighting the beneficial effect of this “synergy” between man and machine.

The Ithaca tool also suggests multiple locations in 84 regions, which can be seen on a map highlighting geographical links across the ancient world.

The tool finally suggests an exact date when the documents were written: 421 BC, about 30 years after the approximate dates suggested by historians.

Today, artificial intelligence contributes to many cultural and artistic fields, such as the restoration of paintings that have undergone many changes with the time factor

"It may seem trivial, but it is essential to our understanding of classical Athens, during the period in which Pericles and Socrates lived," said Thea Somerschild of Ca' Foscari University in Venice, Italy, and one of the study's authors, during a press conference. It is expected that this tool will save thousands of manuscripts and rare books in libraries and museums from damage and restore them for adoption in scientific research and to learn about ancient cultures again.

The researcher pointed out that the “Ithaca” tool can be used in all ancient languages, such as Latin, the ancient Mayan language, or the cuneiform writing of Mesopotamia.

Today, artificial intelligence contributes to many cultural and artistic fields, such as the restoration of art paintings that have undergone many changes with the time factor.

Last year, thanks to artificial intelligence, the famous "Night Patrol" by Dutch painter Rembrandt was recreated more than 300 years after its worst act of vandalism, when the people who transported it decided to cut it up and simply used scissors to cut it on each of the four sides.

Cut out parts were completed in the 18th century and visitors can now see the work in its original size. Thanks to artificial intelligence and a small 17th century copy, scientists have reconstructed the lost pieces that were reprinted and placed around the masterpiece.

Robert Erdmann, a scientist at the Rekimuseum Museum at the head of the restoration project, explains that the solution was to “put artificial intelligence into the art school,” explaining, “To ensure success, I trained three different artificial neural networks to contribute to the process, and this is a type of artificial intelligence that allows us to indoctrinate a device computer by giving him some examples.”

The computer then compared the original and the copy, and, after learning Rembrandt's technique, was turned on to recreate the missing pieces. The simulation was so successful that the machine reproduced even the smallest cracks visible on the work surface, says Erdmann.