AI models make precise copies of cuneiform characters

Deciphering some people's writing can be a major challenge—especially when that writing is cuneiform characters imprinted onto 3,000-year-old tablets.

Now, Middle East scholars can use artificial intelligence (AI) to identify and copy over cuneiform characters from photos of tablets, letting them read complicated scripts with ease.

Along with Egyptian hieroglyphs, cuneiform is one of the oldest known forms of writing, and consists of more than 1,000 unique characters. The appearance of these characters can vary across eras, cultures, geography and even individual writers, making them difficult to interpret. Researchers from Cornell and Tel Aviv University (TAU) have developed an approach called ProtoSnap that "snaps" into place a prototype of a character to fit the individual variations imprinted on a tablet.

With the new approach, they can make an accurate copy of any character and reproduce whole tablets.

"When you go back to the ancient world, there's a huge variability in the character forms," said Hadar Averbuch-Elor, assistant professor of computer science at Cornell Tech and in the Cornell Ann S. Bowers College of Computing and Information Science, who led the research. "Even with the same character, the appearance changes across time, and so it's a very challenging problem to be able to automatically decipher what the character actually means."

Rachel Mikulinsky, a masters student and co-first author from TAU, will present "ProtoSnap: Prototype Alignment for Cuneiform Signs" in April at the International Conference on Learning Representations (ICLR).

An estimated 500,000 cuneiform tablets sit in museums, but only a fraction have been translated and published. "There's an endless amount of 2D scans of these cuneiforms, but the amount of labeled data is very scarce," Averbuch-Elor said.

To see if they could automatically decipher these scans, the team applied a diffusion model—a type of generative AI model often used for computer vision tasks, such as image generation—to calculate the similarity between each pixel in an image of a character on a tablet and a general prototype of the character. Then they aligned the two versions and snapped the template to match the strokes of the actual character.

The snapped characters can also be used to train downstream AI models that perform optical character recognition—essentially turning images of the tablets into machine-readable text. The researchers showed that, when trained with this data, the downstream models perform far better at recognizing cuneiform characters—even ones that are rare or that show a lot of variation—compared to previous efforts using AI.

This advance could help automate the tablet-copying process, saving experts countless hours, and allowing for large-scale comparisons of characters between different times, cities and writers.

"At the base of our research is the aim to increase the ancient sources available to us by tenfold," said co-author Yoram Cohen, professor of archaeology at TAU. "This will allow us, for the first time, the manipulation of big data, leading to new measurable insights about ancient societies—their religion, economy, social and legal life."