Classifying Word-Level Variants in Biblical Hebrew Manuscripts: New Computational Tools for Textual Criticism

The number of variants (or differences) shared by witnesses of the same textual tradition has traditionally been used in stemmatology to determine the witnesses' relationships (Quentin, 1926; Froger, 1968). Importantly, however, the nature of these variants has also been shown to be highly significant (Barthélemy, 1982; Amphoux, 1989).

This presentation will introduce a method developed by our interdisciplinary team for the automatic annotation of variant categories, such as 'plus/minus', 'inversion', 'morphological' and 'lexical'.

The underlying manually annotated data and annotation protocol will first be presented. The different classifier experiments will then be discussed, focusing on the distinction between neural and non-neural models, the development of synthetic data for each variant category, and the use of additional classifier features, such as morphological annotation.

The strongest classifier obtained (F1 score: 0.80) will be presented. The proposed method will briefly be compared with a rule-based alternative.

Finally, future directions will be discussed, including the classifier's integration into a comprehensive pipeline for witness comparison.

Iglika Nikolova-Stoupak's research applies natural-language processing and computational philology to the study of textual transmission. In a 2025 ACL Findings paper with Maxime Amblard, Sophie Robert-Hayek, and Frédérique Rey, she helped develop a classifier for word-level differences in witnesses of Biblical Hebrew manuscripts.

The research classifies aligned word pairs into categories including plus/minus, inversion, morphological, lexical, and unclassifiable variants. The published system combines lexical information, part-of-speech tags, hand-crafted rules, and synthetically derived data and reported an F1 score of 0.80. The research also evaluated neural approaches related to DictaBERT and compared morphological information from manuscript resources.

In 2026, Nikolova-Stoupak and collaborators also presented BEReshiT, an Ancient Hebrew model based on DictaBERT, including a morphological submodel. Together, these projects illustrate how new computational tools can assist stemmatology, manuscript comparison, morphological analysis, and textual criticism while remaining dependent on philological judgment.

Iglika Nikolova-Stoupak is a researcher in computational linguistics and natural-language processing. She works full-time as a research engineer / developer-researcher at LORIA and the University of Lorraine, where a major part of her work applies NLP methods to stemmatology and the analysis of Hebrew manuscripts.

She is also pursuing a Ph.D. in Computational Linguistics at Sorbonne Université. Her doctoral project is titled "Production of Abridged Versions of Literary Texts: A Multilingual Approach."

She holds Master's degrees in Literature from the University of Essex and Computing from Staffordshire University, both completed with distinction, and completed a two-year specialization in natural-language processing at Kyoto University under a MEXT scholarship. In 2025 she received the "Culture and Creativity" award at the Study UK Alumni Awards in France.

Her research combines NLP, historical and low-resource languages, digital humanities, computational philology, multilinguality, textual transmission, and manuscript analysis.