Integrating Semantic and Statistical Features for Authorial Clustering of Qumran Scrolls

Yonatan Lourie, Jonathan Ben-Dov, Roded SharanNAACL 2025

TL;DR

This paper introduces a novel Graph Neural Network (GNN) approach that combines modern AI language models with traditional text analysis to cluster and classify the ancient Dead Sea Scrolls. Their method significantly outperforms existing techniques, offering new insights into long-standing debates about the scrolls' authorship and internal divisions.

What Problem Does This Paper Address?

The Dead Sea Scrolls (DSS) are a treasure trove of ancient Jewish texts, offering unparalleled insights into religious and cultural history. However, their study is fraught with challenges: the scrolls are often fragmentary, making manual analysis laborious and subjective. Furthermore, questions about authorship, sectarian vs. non-sectarian origins, and the internal structure of collections like the Hodayot hymns remain hotly debated among scholars. Traditional methods struggle with the inherent complexities of ancient Hebrew, including its rich morphology, ambiguous word boundaries, and inconsistent orthography. This paper aims to bridge this gap by applying advanced computational techniques to tackle these challenges.

Methods

The core of this research lies in a novel Graph Neural Network (GNN) architecture. The process begins with data collection and preprocessing, where textual content is extracted and cleaned, removing paratextual elements and reconstructed words. Crucially, the text is then chunked into manageable segments.

The innovation lies in how these chunks are represented. The authors combine two types of features:

  • Semantic Features: Modern Hebrew BERT embeddings (specifically BEREL and AlephBERT) are used to capture the deep meaning of the text.
  • Statistical Features: Traditional methods like TF-IDF (Term Frequency-Inverse Document Frequency) and character trigrams are employed to capture word importance and local orthographic/morphological patterns.

These features are then used to construct a graph where text chunks are nodes. Edges are formed based on the cosine similarity between these combined embeddings (TF-IDF and trigram). A Graph Autoencoder (GAE) model, utilizing a two-layer GNN, is then trained on this graph. The GAE learns to encode the graph structure and node features into a low-dimensional latent space, generating refined text embeddings. These embeddings are then used for unsupervised clustering and classification.

Datasets

The primary dataset used is a collection of transcriptions of the Dead Sea Scrolls, sourced from Martin Abegg's prepared data files. This corpus consists of approximately 978 text chunks. For validation, the authors also utilized a dataset of the Hebrew Bible.

Key Insights and Ideas

  • Hybrid Approach is Key: Combining modern deep learning embeddings (BERT) with traditional statistical features (TF-IDF, trigrams) proves more effective than either approach alone for analyzing the unique characteristics of the DSS.
  • GNNs Excel at Fragmentary Text: The graph-based approach is particularly well-suited for handling the fragmented nature of the scrolls, allowing for the capture of relationships between text segments even when they are incomplete.
  • Unsupervised Clustering Reveals Structure: The unsupervised clustering approach successfully identified distinct groupings within the Hodayot hymns, providing strong evidence for the "Teacher Hymns" and "Community Hymns" distinction, and offering new perspectives on transitional passages.
  • Sectarian Divisions Emerge: The model also successfully clustered compositions, aligning with traditional scholarly divisions between sectarian and non-sectarian texts, and even highlighting potential groupings for texts with uncertain identities.
  • Genre Overrides Sectarianism: For rewritten Bible texts, the model tended to group them by genre rather than sectarian affiliation, demonstrating its ability to capture broader textual characteristics.

Key Results

  • The proposed GNN-based method outperforms baseline methods on both the Dead Sea Scrolls and a validation dataset of the Hebrew Bible.
  • The clustering of the Hodayot scroll (1QHª) identified two main clusters corresponding to Teacher Hymns (columns 10-16, with some outliers) and Community Hymns (columns 1-9, 18-23, with transitional chunks).
  • The classification of sectarian scrolls produced two main clusters, with one strongly indicating sectarian content (including Pesharim, MMT, War Scroll, Community Rule) and another containing core sectarian texts like Hodayot and The Rule of the Blessings.
  • The model also accurately identified distinct non-sectarian texts like the Book of Tobit and Book of Jubilees, and captured genre-based groupings for rewritten Bible texts.

Limitations and Future Work

  • Fragmentary Nature: The inherent fragmentation of the DSS remains a significant challenge, and further research into preprocessing techniques could improve results.
  • Subjectivity of Labels: The evaluation relies on predefined labels (sectarian/non-sectarian, authorship) which can be subjective, limiting the absolute objectivity of performance metrics.
  • Pre-trained Models: The lack of a readily available, pre-trained Hebrew model specifically for ancient texts means current models are fine-tuned on more general Hebrew data. Future work could involve creating a dedicated embedding space for DSS texts.
  • Sentence-Transformer Models: The authors suggest that sentence-transformer models could offer greater precision for processing text chunks and are a promising avenue for future research.

Summary updated

BibTeX
@inproceedings{lourie2025integrating,
  author = {Yonatan Lourie and Jonathan Ben-Dov and Roded Sharan},
  title = {{Integrating Semantic and Statistical Features for Authorial Clustering of Qumran Scrolls}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure