Ancient Inscriptions

 

Credit: Acropolis Museum/Socratis Mavrommatis

Ancient civilizations have held fascination for historians and anthropologists ever since humans became curious about the past. Deciphering life in civilizations past meant that excavated tools, vessels, assorted items, inscriptions etc. had to be unearthed and examined. The most informative of these would be the inscriptions on scrolls, papyri, stone, metal, or pottery. A lot of these inscriptions are incomplete due to damage and large chunks of the inscribed text are often illegible. Determining where the texts originated from can be a challenge since they may have been moved multiple times. Dating these texts is another challenge as radiocarbon dating and similar methods can't be used since they can damage the priceless artifacts.

Enter AI.

Researchers from DeepMind collaborated with researchers from the University of Oxford to develop Pythia, an ancient-text restoration system based on Deep Learning. The database used was of the Packard Humanities Institute (PHI) - the largest collection of ancient Greek inscriptions. This collection was converted to machine-actionable text called PHI-ML. This amounted to 35,000 inscriptions, about 3 million words from the 7th to the 5th centuries BCE. Pythia was trained to predict missing letters and words by using words as well as individual characters as inputs.

During prediction, Pythia produced 20 different letters and words that could be filled-in for the missing parts of inscriptions along with confidence probabilities. The historians (domain-experts) would sift through these possibilities and make a final determination.

Ithaca is the latest system which, in addition to text-restoration, makes predictions about the geographical attribution of incomplete inscriptions. The probability distribution over all possible predictions is helpfully visualized on a map. For chronological attribution, Ithaca produces a distribution of its predicted dates between 800 BCE to 800 CE.

Looking at the results, Ithaca is able to produce an accuracy of 62 percent in text-restoration compared to 25 percent for human historians.  But a man-machine combination boosts the accuracy to 72 percent, demonstrating the potential of human-machine collaboration. Ithaca produces location attributions at 71 percent accuracy and date attribution accuracy to within 30 years. These are indeed usable performances.

For more details on this research please go here.