DeepMind claims its AI can decipher ancient Greek texts from damaged artifacts

In 2019, DeepMind, the lab backed by Google parent company Alphabet, announced that it had created an AI system that can restore ancient Greek texts. The lab claimed that the system, called Pythia, could accurately guess sequences of letters in text inscribed on stone tablets that had been cracked, chipped, or otherwise damaged.

Today in a paper published in the journal Nature, DeepMind introduced the successor to Pythia, Ithaca, which the lab says performs even better in Greek text restoration tasks. Ithaca reportedly achieves 62% accuracy in restoring damaged texts, 71% accuracy in identifying their original location, and can date texts to within 30 years of their date ranges.

DeepMind partnered with Google Cloud and Google Arts & Culture, Google’s cultural preservation nonprofit, to launch an interactive version of Ithaca. It also open-sourced the code and the model that powers the system.

Roger Bagnall, a professor of history at New York University, is hopeful that Ithaca can be extended to other ancient languages, particularly those for which few examples exist. “The dynamism of Ithaca is particularly appealing; looking at the improvement in performance since Pythia gives hope that even the excellent results of Ithaca can before long be improved, with iterative learning based on the human-machine collaboration that it makes possible,” he said in a statement.

Speaking Greek

Ithaca is a collaboration between DeepMind and the Department of Humanities of Ca’ Foscari University of Venice, the Classics Faculty of the University of Oxford, and the Department of Informatics of the Athens University of Economics and Business. The goal was to build a system that can decipher Greek text written on stone, pottery, and metal artifacts, some of which dates back to over 2,500 years ago.

The challenge is twofold: ancient Greek inscriptions are often damaged and modern dating techniques, like radiocarbon dating, can’t be used.

Building on its work with Pythia, DeepMind developed Ithaca using a dataset of over 178,000 Greek inscriptions supplied by the Packard Humanities Institute. Researchers at the lab trained the system using Greek words and individual characters, so that damaged or missing text wouldn’t interfere with Ithaca’s ability to analyze either.

It’s different from the approach typically taken with text-analyzing and -generating systems like OpenAI’s GPT-3, which are trained using only sequences of words. The order in which the words appear in sentences and the relationships between them provide extra meaning and context to the systems. Ithaca had to learn to make do without this information.

In an illustration of how Ithaca might be useful to historians, DeepMind says that the system predicted a date of 421 BCE for a series of Athenian decrees — for example, awards of citizenship, declarations of war, and enactments of treaties — made at a time when notable figures such as Socrates and Pericles lived. Originally thought to have been written before 446/445 BCE, the system agreed with new evidence that suggests a date of the 420s BCE.

DeepMind says it’s working on versions of Ithaca trained in other ancient languages. In the meantime, historians can use datasets in the current architecture to study other ancient writing systems, the lab notes — including Akkadian, Demotic, Hebrew, and Mayan.

“Ithaca’s extensibility to other languages and textual corpora is exciting. I can hardly wait to see it applied to the documentary papyri, where we have far more precise dating but far more unprovenanced texts, because of the operations of the antiquities market,” Bagnall continued. “It should be possible with Ithaca’s help to reconstruct the workings of that market and the original historical context of many more of the thousands of papyrus documents.”

Restoring texts with AI

DeepMind isn’t the first to apply AI to historical texts. Increasingly, academics have been exploring machine learning to restore documents that were previously lost to history, including those written in cuneiform.

For example, last year, researchers at Jerusalem’s Hebrew University created an AI system that can predict missing words, phrases, and sentences from cuneiform tablets up to 4,500 years old. Elsewhere, a team of researchers in Italy used a robotic system to process, match, and physically reconstruct frescoes and other shattered artifacts from Pompeii.

But AI designed for artifact restoration raises questions about whether the process could influence or change the meaning of the original work. After all, AI, like humans, isn’t infallible — Ithaca made errors in restoring damaged text 38% of the time.

DeepMind’s solution is visual aids aimed at minimizing the potential for misinterpretation of Ithaca’s predictions. Ithaca offers several text restorations “hypotheses” from which users can choose, each with a different associated confidence metric. The system returns probabilities for 84 different ancient regions, representing its level of uncertainty. Ithaca also produces a distribution of predicted dates across decades from 800 BCE to 800 CE, with a confidence value for specific ranges, and highlights words that led to its predictions for text, location, and dates.

Alison Cooley, president of the international digital epigraphy association and a professor at the University of Warwick, doesn’t believe that systems like Ithaca will replace the need for human expertise. Instead, they can act as a guide or tool for researchers studying antiquities, he says — perhaps helping to uncover patterns that’d otherwise be missed. In a DeepMind experiment, expert historians were 25% accurate in restoring ancient texts, but their performance increased to 72% when using Ithaca.

“This paper represents a very important development in the collaborative use of AI to enhance the restoration, dating, and attribution of inscriptions written in Greek from the ancient world over a period of several centuries,” Cooley said in a statement. “The innovative design of Ithaca promises to transform the potential contribution of inscribed evidence to our understanding of key moments in world history.”