AI Tackles Ancient Forgotten Languages: What Neural Networks Can Do

Ancient inscriptions often consist of mysterious symbols on stone or clay, representing cultures that have vanished. Decipherment is challenging without a Rosetta Stone or related languages. Neural networks have recently been applied to restore missing portions of texts, identify patterns across thousands of inscriptions, test hypotheses, and estimate age and origin. They are particularly good at finding repeated sequences and filling gaps, using both text and image. They can also search for correspondences between related languages, mass-check assumptions, and provide dating and geographic attribution. For example, Simon Cordwell used a program to test a hypothesis about Linear A, proposing readings for 40 signs and a 408-word dictionary. The Vesuvius Challenge team used tomography and neural nets to virtually unroll a carbonized scroll, revealing a philosophical treatise. Google DeepMind's Aeneas model for Latin inscriptions achieves 73% accuracy in filling gaps, 72% in province attribution, and errors of about 13 years in dating. Ithaca, for ancient Greek, restores texts with 62% accuracy, and when used by historians, their accuracy improved from 25% to 72%. For Chinese oracle bones, generative adversarial networks suggest potential missing characters. However, neural networks cannot deduce what is not in the data; if too few texts exist, they can only offer plausible but unverified hypotheses. They predict likely continuations but do not understand the meaning. Therefore, they remain tools for researchers, speeding up hypothesis testing and discovery, but cannot fully decipher a forgotten language on their own.