Recent developments in the computational analysis of the Bible

By Samuel Swire and Thijs Amersfoort

In his article Computational Linguistic Analysis of the Biblical Text Willem van Peursen gave an overview of the development of the computational analysis of the Bible from its early beginnings through 2022. Since then, the development within Artificial Intelligence (AI), especially Natural Language Processing (NLP), has continued to grow. In this blogpost we highlight some of the major advancements to-date within these fields as they apply to Digital Humanities (DH) broadly and Biblical Studies in particular.  

One of the main drivers of AI innovation comes from transformer architectures, in particular for large language models (LLMs). Transformers were first announced by researchers at Google in 2017¹, which subsequently gave rise to the advent of other architectures such as Bidirectional Encoder Representations from Transformers (BERT)² in 2018. Concurrently, OpenAI published a paper outlining methods for enhancing such models, resulting in the release of the first Generative Pre-trained Transformer (GPT).³  GPT was one of the first models that could generate natural language responses to natural language prompts. When trained on large corpora of text and combined with reinforcement learning, these models excel at understanding and generating human language, making them highly capable at tasks including translation, answering questions about a text, summarizing, sentiment analysis, entity recognition, and topic modeling.  

Transformer-based models and LLMs have led to many studies and new applications in DH. To begin with, the University of Leiden has published two LLMs based on BERT, trained on historical Dutch and English corpora.⁴ These models are designed to facilitate the interpretation of historical texts, for which LLMs trained on modern languages may perform poorly. In a similar application, OpenAI’s GPT-4 was employed to translate and analyze a substantial corpus of travelogues as part of the DEHisRe project.⁵ The researchers observed that the initial outcomes of utilizing these LLM were encouraging, with the models exhibiting “superior flexibility over traditional machine learning methods”. However, they also identified potential limitations. The researchers highlight the substantial resources required for the operation of these models and their propensity to “hallucinate,” which introduces risks to the reliability, reproducibility, and factual integrity of the outcomes.⁶ 

The application of within the Hebrew Bible has likewise increased rapidly. In paleography and codicology, for instance, the usage of transformer-based with unsupervised learning techniques^7,8,9 has aided in classifying DSS fragments and documents within the Cairo Genizah.^10,11,12,13 Within textual analysis, the application of more advanced statistical and NLP models has become more common, as has the awareness of the limitations of these models on the Bible due to the limited corpus size.^14,15 Researchers have begun to utilize data enrichment and corpus enlargement techniques such as constructing datasets and training models with closely related Semitic languages (e.g. Hebrew and Syriac)¹⁶, as well as focusing the tuning stages on low-shot or zero-shot approaches.¹⁷ The application of LLMs has also led to methodological reflections on the limitations of models pre-trained on modern corpora and/or non-Semitic language corpora in performing tasks on classical texts like the Hebrew Bible.^18,19   

Given the concerns surrounding reliability, reproducibility, and factual integrity, the outputs generated by such models will invariably necessitate verification. It remains to be seen whether this will improve with future advancements, as well as what novel applications will emerge in the field of humanities.