Abstract

The study of ancient writing systems poses unique challenges for any deep learning method. If we want to apply these technologies to an undeciphered writing system, we need to use an unsupervised approach, since no source of supervision can be derived from data. The application of these methods, however, requires careful considerations about the nature of the writing system and it is necessary to have some form of validation for the results. Therefore, the obvious choice is to use a deciphered writing system to validate and develop our methods, then move to an undeciphered script. The framework that ensues is a good example on how to deal with this situation, adapting to the peculiarities of each writing system we investigate. While the specifics cannot be generalized, the approach can be applied to other scripts as well.