Restoring, Placing, and Dating Ancient Texts with the Help of AI and Historians
The birth of human writing marked the beginning of recorded history and is vital to our understanding of the past and our present world. Ancient Greek inscriptions provide valuable insights into the Mediterranean region, covering topics from laws and calendars to oracles and leases. However, many of these inscriptions have been damaged or displaced over time, making interpretation challenging. That’s where AI comes in.
DeepMind, in collaboration with Ca’ Foscari University of Venice, the University of Oxford, and the Athens University of Economics and Business, has developed Ithaca, a deep neural network designed to restore damaged inscriptions, determine their original location, and establish their date of creation. This groundbreaking technology opens doors for collaboration between AI and historians, enriching our understanding of ancient history.
Ithaca achieves a 62% accuracy in restoring damaged texts, a 71% accuracy in identifying their original location, and can provide a date range within 30 years of the inscription’s true age. Historians have already used Ithaca to reassess significant periods in Greek history, shedding new light on the past.
To make Ithaca accessible to researchers and educators, it has been launched as a free interactive tool on Google Cloud and Google Arts & Culture. Additionally, the code, pretrained model, and an interactive Colaboratory notebook have been open-sourced for further research and development.
Ithaca’s training dataset consists of a vast collection of Greek inscriptions from the Packard Humanities Institute. To ensure accuracy, the model was trained using both words and individual characters as inputs. By evaluating these inputs in parallel, Ithaca effectively analyzes damaged and incomplete inscriptions.
The results provided by Ithaca are also designed to be easily interpretable by historians. Restoration hypotheses allow experts to choose from multiple predictions generated by Ithaca, while geographical attribution provides a probability distribution of possible locations. For dating inscriptions, Ithaca produces a distribution of predicted dates, enabling historians to gauge the model’s confidence. Saliency maps highlight the most influential words in Ithaca’s predictions, emphasising contextually important information.
The effectiveness of Ithaca as a research tool is evident in our evaluation. Historians achieved a 72% accuracy in restoring texts when using Ithaca, surpassing their individual performance of 25%. This collaboration between humans and machines shows promise in advancing historical interpretation and contributing to ongoing debates within the field.
One example is the dating of Athenian decrees from the time of Socrates and Pericles. While previous estimates placed these decrees before 446/445 BCE, Ithaca’s analysis aligns them with new evidence suggesting a date in the 420s BCE. This highlights how AI can contribute to revising our understanding of significant historical moments.
Our vision extends beyond Ancient Greece, as we are working on versions of Ithaca trained on other ancient languages. Historians can already utilize Ithaca’s current architecture to study various writing systems, from Akkadian to Demotic and Hebrew to Mayan. By embracing AI, we hope to revolutionize the study and documentation of major periods in human history, fostering collaboration between AI and the humanities.