Master's Thesis : Papyrus matching using machine learning

Abstract

For papyrologists, the task of reconstructing papyri from scattered fragments is a tedious and time-consuming, yet crucial task to uncover new texts that could provide meaningful information on a variety of subjects. With the advent of digital humanities, several tools have been developed to make papyrological studies more effective, in particular when it comes to the digitization and manipulation of papyri. But none of these tools can help papyrologists for the decision-making involved in the reconstruction of papyri. In recent years, it has been shown that deep learning techniques can be successfully applied to a wide variety of problems, sometimes with impressive results. However, the application of such techniques to papyrology is pretty much novel.

In this thesis, deep learning techniques are applied to the task of papyrus fragment matching, to evaluate the possibilities offered by neural networks to tackle this complex task. Using a papyrus fragment dataset built from scratch using raw data from the papyrus collection of the Museo Egizio of Turin, several siamese neural network architectures are evaluated and compared. The best performing model is also evaluated on more a realistic use case and gives some encouraging results, though it is not yet ready for practical use. Several improvements are proposed to further improve the results.