ResponsaNet:

Construction and analysis of the citation networks in the Responsa literature

Led by: Prof. Zhitomirsky-Geffet, Dr. Schler, Dr. Katzoff and Mr. Ben-Gigi

Funded by the Ministry of Science and Technology

​​

​Rabbinical literature is characterized by a multiplicity of viewpoints and diverse and contesting opinions rooted in earlier sources whose influence persists in subsequent essays. One of the challenges in studying literature of this kind is to identify and organize the many controversies and views and to examine their development over the ages and generations. This project offers a new paradigm and conceptual framework for the study of viewpoint plurality in historical literature through the prism of citation networks that are embedded in it, as well as a computational methodology based on advanced machine learning algorithms that will be applied in the subfield of rabbinical literature as a case study. Firstly, an ontology of citation data will be semi-automatically constructed based on the Bar-Ilan Responsa Project texts which will then be utilized to generate directed bipartite citation networks (where authors and books will constitute the nodes and citation relationships between them will represent the edges) at different levels of specification, e.g. the global network for the entire corpus, local networks for selected subsets of books or groups of authors, and personal networks for specific books and authors. Next, we will apply methods and knowledge used in traditional rabbinical literature research to investigate whether, how and to what extent the viewpoints of the authors are reflected in the obtained networks, and whether authors with strong relationships in the networks share the same views. Although the study will examine the proposed paradigm in the Responsa Project corpus, the proposed methodological framework can be applied to the entire Jewish Bookshelf, as well as other multi-viewpoint corpora of literature in other languages, genres and periods. The developed technology will be incorporated in the Bar-Ilan Responsa project and will ensure its long-term sustainability and powerful impact on scholars and learners. The project is funded by the Israeli Ministry of Science and Technology.

 

In this study, we used a subset of the Responsa literature written between the years 1000 and 1500 CE. This period is called in Rabbinic literature the period of the Rishonim. The corpus contains about 30 books (i.e., collections of Responsa works where almost every single collection was written by a single author), that sums up to a total of 12,193 Responsa works (i.e., individual answers to Halachic questions). The size of the Responsa book varies between dozens and hundreds of answers. Some answers are quite short and contain only a few lines, while others are thousands of words long. The corpus that covers all the known Responsa literature from the period was supplied by the Bar-Ilan Responsa Project. Our experiments showed that reference extraction is a complex task that cannot be effectively implemented by a single learning model. Therefore, based on the above generic models, two separate task-specific sub-models were trained for the following tasks: 1) a sub-model for the identification of atomic entities in the reference; 2) a sub-model for the identification of the start/end positions of the reference in the text.

The best results have been achieved by BERT-CRF that was pre-trained on Rabbinic Hebrew texts (with precision 0.896, recall 0.905, and F1 score is 0.901) as shown in Figure 1. The lowest performance was obtained by the BERT model trained on modern Hebrew, due to the various characteristics posed by historical Hebrew texts that do not exist in the modern Hebrew language.