Using deep neural network models to aid in the decipherment of Linear A

Linear A is a script that was used on Crete and its surrounding islands by the Minoan civilization during the Bronze Age (ca. 1800–1450 BCE). It was first discovered in 1900 AD at the site of Knossos, alongside another related, but slightly later script called Linear B, which was associated with Mycenaean culture and attested on both Crete and mainland Greece (ca. 1400–1200 BC).

While Linear B was shown to represent an early form of ancient Greek in 1952 AD, Linear A remains undeciphered and appears to encode an unrelated language. Attempts at decipherment using conventional, manual methods have proven unsuccessful to date, owing in no small part to the restricted size of the current corpus of texts.

However, in recent years, deep learning techniques have shown great success as a tool in deciphering other ancient scripts. By combining expertise in Aegean scripts from the University of Melbourne’s Classics and Archaeology department with the technical capabilities of the MDAP team, we are hoping to investigate the nature of Linear A through language models trained via deep learning, working towards an effort at machine decipherment.

A key part of this project will involve pre-training models in potentially related languages, before refining the models on Linear A, to test the level of cognancy between them. The identification of a cognate language for Linear A would lay the groundwork for machine decipherment – a method already successfully applied to other ancient languages with known cognates, including Ugaritic and Linear B.