TL;DR
This paper proposes a novel approach to dating biblical texts by integrating traditional biblical scholarship with machine learning, specifically Large Language Models (LLMs). By combining redaction criticism and linguistic analysis, the authors aim to uncover the complex compositional history of the Book of Jeremiah, a notoriously difficult task due to its long and debated history of editing.
What Problem Does This Paper Address?
Dating biblical texts, particularly the Hebrew Bible, is a significant challenge. While we know these texts were composed within the first millennium BCE, pinpointing their exact dates is often debated. The Book of Jeremiah, being one of the longest and most complex, exemplifies this problem. It has a long history of composition and editing, with unclear historical backgrounds for many of its sections. Traditionally, biblical scholarship has relied on distinct fields like redaction criticism (analyzing layers of editing) and historical linguistics (studying language evolution). However, these fields often operate separately and can yield conflicting results. This paper aims to bridge this gap by developing an integrated methodology that leverages the power of machine learning to combine these approaches.
Methods
The core of the proposed methodology involves using Large Language Models (LLMs) for linguistic analysis. The authors plan to train an LLM on the Hebrew Bible, specifically addressing the challenge of it being a "low-resource language" in NLP terms. They will explore different architectural choices, text representations (vocalized vs. unvocalized), and tokenization strategies. A key aspect of their approach is the use of transfer learning, potentially training on related languages to improve performance. After training a masked language model, they will fine-tune it to act as a text classifier. This classifier will be used to segment and date parts of the Book of Jeremiah. Crucially, the authors emphasize the importance of "explainability" in their LLM approach, ensuring that the models' predictions are meaningful from a linguistic perspective, rather than being black boxes.
Datasets
The primary dataset for this research is the ETCBC dataset of the Hebrew Bible. The authors also mention using labeled data drawn from "Classical Biblical Hebrew" (CBH) and "Late Biblical Hebrew" (LBH) texts for initial experiments. They also consider using texts in related languages for transfer learning.
Key Insights and Ideas
- Integration is Key: The paper argues that combining redaction criticism with linguistic analysis, facilitated by LLMs, is crucial for a more robust understanding of biblical text composition.
- LLMs for Low-Resource Languages: The research tackles the challenge of applying powerful LLMs to a language like Biblical Hebrew, which has a relatively small corpus compared to modern languages.
- Explainability Matters: The authors highlight that simply getting accurate predictions from LLMs is not enough; understanding how the models arrive at their conclusions is vital for linguistic insights.
- Jeremiah as a Test Case: The Book of Jeremiah is chosen as a complex and representative example to demonstrate the effectiveness of their integrated methodology.
Key Results
The paper presents initial experimental results from Wilson-Wright, who finetuned a RoBERTa Base model (named "BERIT") for distinguishing between different linguistic phases of Biblical Hebrew. This model achieved 73.4% accuracy on a validation dataset. An explainability framework revealed that the model learned at least one feature that scholars have identified as distinguishing CBH from LBH: the occasional spelling of the personal name "David" as דָּוִיד in LBH, versus דָּוִד elsewhere. Another experiment by Naaijer used an LSTM-based sequence classifier to distinguish between CBH and LBH, finding that the books of Jonah and Ruth share more characteristics with CBH. However, the authors note that LSTM models are considered black boxes.
Limitations and Future Work
The authors acknowledge that dating biblical texts using language alone is a complex task, and some scholars have contested its feasibility. They also note that current LLMs can be "black boxes," making it difficult to understand their prediction mechanisms.
Future work will focus on:
- Determining the optimal architecture and text representation for training an LLM on Biblical Hebrew.
- Exploring transfer learning from related languages.
- Finetuning the LLM to classify parts of the Book of Jeremiah.
- Ensuring the explainability of the LLM's predictions is a central part of the research process.
- The ultimate goal is to develop a comprehensive methodology that integrates redaction criticism and linguistic analysis to shed light on the compositional history of the Book of Jeremiah.
Summary updated