Introduction
This poster addresses a collaboration between ancient historians and computer scientists which uses machine learning to analyse the planning processes involved in creating Latin inscriptions. It represents a pilot project funded by an Institute for Data Science and Artificial Intelligence Research Award at the University of Exeter, led by Dr. Charlotte Tupman, Dr. Jacq Christmas and Dr. Dmitry Kangin.
This contribution addresses the theme of the open data movement: its entire dataset comprises materials made available by the Epigraphische Datenbank Heidelberg (EDH) as open data. This project could not have taken place without the principles of open data being embraced by the digital classicist community and the EDH team.
Aims
Inscriptions are one of our most important sources for the ancient world, covering a vast number of subjects, monument types, geographical areas and periods [2]. There were several stages in their planning: the preparation of the surface, drafting of the layout of the text, possibly the filling in of letters with brush strokes, and finally the carving [3, 4]. Our interest is in the drafting: were Roman craftsmen in different regions working to defined sets of planning ‘modules’ which could be scaled up or down? Did the stonecutter have freedom to influence the design after drafting? Does this vary by time or place? Could we even begin to identify the work of specific workshops [5]?
Methodology
The project applies a neural network, investigating Neural ODEs [6] and extending the model for this image processing case. We investigate whether the model can learn effectively, and address the development of a new basis for training Neural ODEs using a mathematical method, invariant embedding. Once characters in an image have been localised, we analyse the regularity of their size, shape, spacing, position and orientation, and the overall shape of their outline.
Summary
Improving our understanding of the planning out of inscriptions is important not only from the perspective of ancient craftsmanship. Inscriptions are frequently fragmentary, and epigraphers have to determine what letters would have been present on the missing sections. Reconstructions of inscriptions influence our interpretation of ancient sites, so being able to predict accurately the positioning of the letters is crucial. We seek to initiate valuable conversations with DH specialists in other fields which may influence the future direction of our investigations.
References
[1] Epigraphic Database Heidelberg, https://edh-www.adw.uni-heidelberg.de/
[2] J. Bodel (ed.), 2001. Epigraphic Evidence: Ancient History from Inscriptions, London, p. 4
[3] R. Grasby, 2009. Processes in the Making of Roman Inscriptions. Study 11, p.12
[4] G. Susini, 1965. The Roman Stonecutter, Oxford
[5] B. Russell, 2013. The Economics of the Roman Stone Trade, Oxford
[6] T.Q. Chen, Y. Rubanova, J. Bettencourt and D.K. Duvenaud, 2018. Neural Ordinary Differential Equations. In Advances in Neural Information Processing Systems, pp. 6571-6583