Using Large Language Models for Interactive Exploration of Historical Jewish Texts and Rabbinical Literature

• In Event: Digital Innovations and the State of the Field

Abstract

This report outlines the successful implementation on how to interact with a Large Language Model (LLM) within historic Jewish texts. The texts were taken from articles on rabbinical literature within Compact Memory, a collection of European Jewish journals and periodicals spanning from 1778 to 2006. Source texts from rabbinical literature included German translations by Hoffmann and Hebrew/Aramaic versions from Sefaria.

The aim was to identify direct quotes, indirect quotes, and paraphrases from rabbinical literature within Compact Memory with the use of a Large-Language-Model. But the result was that the LLM can not only assist scholars in analyzing and extraction information form historical documents even in a different language and interact and discuss the content.

The challenge, serving as a Proof-of-Concept aimed to demonstrate the feasibility of utilizing LLMs as a research tool for scholars. Through interaction with the LLM, users could search for specific quotes, request summaries and translations. The future goal of this approach is to generate networks of text reuse within rabbinical literature.

The methodology employed automated querying of individual articles from Compact Memory, alongside with manual evaluation and verification of extracted quotes. Challenges inherent in this process included the varying quality of Optical Character Recognition (OCR), layout segmentation, historic language and the multilingual nature of source texts, requiring automatic translation between German and Hebrew/Aramaic. Despite these challenges, the Large-Language-Model was able to identify direct quotes, indirect quotes and paraphrases from rabbinical literature within Compact Memory Journals.

The study underscores the potential of LLMs as research tools for scholars engaging with historic texts. By automating certain aspects of information retrieval and analysis, LLMs streamline the research process and enable scholars to explore vast collections of historical documents more effectively.

In summary, the challenge achieved its goals, demonstrating the potential of using LLMs for quote identification within historical journals. The outcome of this challenge paves the way for further exploration of LLMs in historical research. Future endeavors may focus on refining the interaction between scholars and LLMs, improving OCR quality, exploring additional functionalities to enhance the utility of LLMs or even training a custom Model.

Author

• Benjamin Schnabel, University of Applied Sciences, Mannheim