DeepMind AI Outperforms Scholars in Restoring Ancient Greek Inscriptions
DeepMind introduces Pythia, an AI system that accurately predicts missing text on damaged ancient Greek stone tablets. The machine learning tool achieves a 70 percent accuracy rate, significantly surpassing the efforts of human PhD students.
DeepMind creates an AI system called Pythia that helps scholars restore fragmentary ancient Greek texts on broken stone tablets. These inscriptions, which date back as much as 2,700 years, serve as invaluable primary sources for history and literature, but they often contain missing pieces known as lacunae that make translation incredibly difficult.
The researchers build a complex pipeline to convert the world's largest digital collection of ancient Greek inscriptions into machine-readable text. This allows the algorithm to learn patterns and accurately guess missing sequences of letters in a manner similar to how a person completes a familiar sentence.
In direct tests, Pythia achieves an impressive 70 percent accuracy rate in filling in artificially removed text, while human PhD students only reach about 43 percent accuracy. Furthermore, the correct interpretation appears in the AI's top 20 suggestions 73 percent of the time, proving it as a powerful new tool for the painstaking field of epigraphy.