Abstract

Computational methods and machine learning have been applied to problems in the study of Proto-Elamite writing in recent years, with results which confirm that the available digital format of the corpus and the work-in-progress signlist can support meaningful computational analysis. However, the specialist’s system for transliterating the texts nonetheless determines the ways that algorithms have been learning sign relationships. This paper therefore explores how image-aware language modeling can address difficulties arising from biases introduced into the data during transliteration. This approach can help us to better understand the system of signs and to revise the proto-Elamite signlist, a fundamental task of the decipherment project.