Prediction of Paleographical Features in Ashkenazi Square Hebrew Script for Identifying Subclusters Within the Type-mode

• In Event: Computational Approaches

Abstract

This paper presents a model for the paleographical features of document images written in Ashkenazi square Hebrew script, structured hierarchically with multi-labels and mutually exclusive sub-labels. We train a convolutional neural network (CNN) to predict these labels and use the predictions as paleographical feature vectors for each document image. These vectors serve as the basis for clustering the document images, thereby unveiling hidden subgroups within the Ashkenazi square script style.

The study employs a curated dataset from 55 manuscripts, each contributing four pages. These manuscripts lack specific annotations about date or region but include distinct groups believed to be from France and Germany, as well as unique manuscripts from England prior to 1290. This approach addresses the limitations of conventional algorithms, such as the bag of words method, which was originally developed for natural scene objects that possess a consistent number of vital features - a condition not applicable to the paleographical features of document images. These traditional methods have been relatively ineffective in identifying the critical features necessary for paleographical clustering. By employing a deterministic methodology, as guided by expert paleographers, and executing a brute-force search to optimize cluster formations, this study not only identifies subclusters but also enhances the overall understanding of sub-clustering, with potential applications to other lesser studied Medieval Hebrew script types like Byzantine and Yemenite.

This systematic method addresses the challenge paleographers face in simultaneously remembering and analyzing features across multiple pages to discern grouping patterns, providing well-defined clusters, and revealing driving features of these formations. The ability to automatically cluster script type-modes will be very useful for human paleography and will dramatically advance computational paleography for Hebrew and other languages.

The research is performed at Tel Aviv University; the team includes Dr. Berat Kurar-Barakat, Dr. Daria Vasyutinsky Shapira, Mr. Sharva Gogawale, Mr. Mohammad Suliman, and Prof. Nachum Dershowitz.

Funded by the European Union (ERC, MiDRASH, Project No. 101071829). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.

Authors

• Daria Vasyutinsky Shapira, Tel Aviv University

• Nachum Dershowitz