We propose an AI-powered platform to assist experts in transcribing, dating, identifying, and editing ancient manuscripts. In this paper, we discuss our ongoing work on AI-assisted Greek papyrology and our vision for a broader application that is intuitive for scholars of the ancient world. We envision this platform as an all-in-one system for AI-assisted papyrology that can be extended to additional languages and media.
In 2011, a Zooniverse.org collaboration called the Ancient Lives project began crowdsourcing the transcription of papyrus fragments housed at the University of Oxford, such as the one shown in Figure 1. The project resulted in millions of annotations. These highly damaged fragments are challenging for most modern handwritten text recognition (HTR) methods.
The AL-ALL dataset, which was derived from the crowdsourced annotations collected during
the Ancient Lives project, consists of 419,445 Greek characters, representing all 24 characters of
the Greek alphabet, cropped from images of papyrus fragments. Due to ongoing papyrological
research, only 205,797 character images from published papyri were made available as the
AL-PUB dataset, shown in Figure 2 and available at https://data.cs.mtsu.edu/al-pub/. As
demonstrated in Swindall et al. 2021 [
One of the greatest challenges in crowdsourcing datasets is sampling bias. This was especially
true for the AL-ALL and AL-PUB datasets. In Swindall et al. 2022 [
Utilizing models trained on AL-ALL, several machine learning tools have been developed that
form a handwritten text recognition (HTR) pipeline. This pipeline expedites the process of
producing a diplomatic transcription, which constitutes an un-edited typescript of the text
visible in a given manuscript. The first tool, a character segmentation model, used transfer
learning to re-task YOLOv5 with locating characters within papyrus images. The second tool
is a character classification model, with a validation accuracy over 94%, which is based on a
ResNet architecture used in Swindall et al. 2021 [
A routine task in papyrology is accurately dating manuscripts. In the case of documentary papyri (accounts, letters, leases, etc.), the scribe usually dates the manuscript; though the date is often lost due to damage. Literary papyri (ancient books) never contain a date, unless portions of it were reused for documentary purposes. In the absence of a date, papyrologists must infer it by comparing the handwriting with other dated manuscripts. To automate this process, a pipeline of models was created that can classify a fragment according to classes representing a period of two centuries (i.e., 400 BCE - 201 BCE, 200 BCE - 1 BCE, etc.), with a range of dates spanning from 400 BCE to 600 CE. To create this pipeline, images of documentary papyri with known dates were run through our HTR pipeline, thus obtaining dating information at the level of individual characters. Models were then trained via transfer learning on the ResNet classification model to attribute dates to individual characters. However, due to the high variability of handwriting styles, individual character dates can be unreliable. To address this, a Gaussian Process model was created which assigns a date to an entire fragment based on the predicted dates of its constituent characters. When trained on fragments with 25 or more characters, this model achieves a precision and recall of 75%-80%. Currently, we are investigating possible ways of increasing the temporal resolution without diminishing prediction quality.
Digital Epigraphy, which produced digital editions of ancient inscriptions, continues to be a
promising area of natural language processing (NLP) research. Eforts such as Pythia [
Beyond textual reconstruction, we believe it may be possible to use computational and
deep learning methods for tasks including document identification, provenance, and detection
of classification errors for existing digital editions. For example, Williams et al. 2014 [
With the increasing development of AI tools to assist in ancient manuscript research, it will
be necessary to modernize the existing infrastructure for creating and managing born-digital
editions of ancient manuscripts. Our Proteus platform, Williams et al. 2015[
Currently a solution is in development which will utilize blockchain and smart contract
technologies for the management and storage of digital editions. In this proposed system,
illustrated in Figure 4, smart contracts are created for new, original editions. Rather than storing
all data in complex databases, this smart contract stores only the location of the data on the
blockchain itself. Editors of critical editions can submit their edition to the smart contract, which
then stores the location of the critical edition’s data. The editions themselves can be stored in a
number of ways: on a local server where the blockchain is hosted, on a public blockchain, or
in a distributed file storage platform such as the InterPlanetary File System (https://ipfs.tech/).
Beyond ofering a less complex method of edition management, blockchain and smart contracts
ofer an avenue to a more transparent and decentralized peer-review ecosystem, as discussed in
Tenorio-Fornés et al. 2021 [
Although we have developed a suite of AI-enabled methods to study the papyrology as a proof of concept application, these tools remain out of reach for many scholars in the field. We envision the creation of a holistic platform which incorporates a host of tools that assist in transcribing, dating, identifying, and editing manuscripts. Figure 3 shows an example transcription. Our approach is likely transferable to other kinds of manuscripts and languages. Instead of a platform limited to Greek papyrology, we envision one that can be interoperable with other language and manuscript datasets from the ancient world.