Artificial intelligence tool helps restore and attribute ancient texts.
Ithaca is a deep neural network that, while on its own achieves 62% accuracy in repairing damaged texts, is intended to be used collaboratively with historians.
The inscriptions, texts written on durable materials, are a very good source to learn about the life and society of the times. One of the main problems historians encounter is that these inscriptions are sometimes incomplete or have been moved from the place where they were created. In 2017, Yannis Assael, artificial intelligence scientist and current Google DeepMind researcher, and Thea Sommerschield, historian and Marie Curie fellow at the University of Venice, discussed these challenging tasks for historians and concluded the great cooperative potential between its disciplines. This is how Ithaca was born, a deep neural network for textual restoration and geographic and chronological attribution of ancient Greek inscriptions.
This tool is designed to work together with history professionals: by itself it achieves an accuracy of 63% to restore damaged texts, but together with the work of historians it manages to improve the accuracy of professionals from 26% to 73%. . The percentage of accuracy in the location is 71% and, in addition, it can date the inscriptions in less than 30 years. In order to maximize collaboration between historians and deep learning, this tool offers multiple theories. For restoration, Ithaca provides 20 predictions decoded and ranked by probability. In this way, it is easier for historians to choose between the suggestions of the tool, taking into account their knowledge. Regarding geographic attribution, the instrument classifies the results among 84 regions; the list of candidate regions is implemented in a map and bar chart.
Finally, instead of giving a value for chronology, it predicts a categorical distribution over dates, which are grouped into 10-year intervals, between 800 BC and 800 AD. Ithaca is a type of artificial intelligence called a deep learning model. It is based on a neural network, inspired by the neural networks found in the human brain, according to the researchers. “We've trained computers to use these neural networks for large amounts of data, so they can apply what they've learned to new data they haven't seen before. It is trained on the largest dataset of ancient Greek inscriptions. Another aspect that stands out is that, due to the architecture of the tool, it is easily applicable to any ancient language, “from Latin to Mayan and Akkadian”, in addition to any written medium.
