Introduction
Medieval legal scholarship relied on a sophisticated system of cross-references long before the advent of digital hyperlinks. In foundational texts such as the Liber Extra (the Decretales Gregorii IX), the primary legal code is enveloped by the Glossa Ordinaria—a dense layer of commentary containing thousands of allegationes, or legal references. These citations point to other laws, papal decrees, or Roman law principles, forming a vast intellectual network.
However, these references are notoriously difficult to process: they are highly abbreviated, non-standardized, and embedded in Latin text devoid of modern punctuation. Manually indexing these "links" is a monumental task for historians. This research presents a high-accuracy pipeline designed to automatically identify and index these references, enabling digital navigation of medieval legal thought.
Methodology
The researchers framed the identification of allegationes as a Named Entity Recognition (NER) task, utilizing the BILOU (Beginning, Inside, Last, Outside, Unique) labeling schema. The workflow transitioned from manual expert annotation to large-scale automated processing.
The study compared two primary machine learning approaches:
- Conditional Random Fields (CRFs): A statistical sequence modeling approach. The team tested a "basic" version and a "rich" version that incorporated bigrams and trigrams within a broader context window.
- Fine-tuned Large Language Models (LLMs): The researchers evaluated standard BERT and Latin BERT, the latter being pre-trained specifically on classical and medieval Latin corpora.
The dataset was derived from a digital transposition of the 1582 Editio Romana of the Liber Extra. A domain expert manually annotated 12 randomly sampled chapters (titles) using the INCEpTION platform to create a gold standard for training and validation.
Results and Performance
The experimental results revealed a surprising victory for "traditional" machine learning over modern deep learning architectures. The "Rich" CRF model achieved a near-perfect accuracy of 0.978, significantly outperforming Latin BERT (0.924).
Several key factors contributed to these results:
- Tokenization Challenges: Standard BERT performed poorly (0.837) because its tokenizer struggled with Latin morphology, often fragmenting words into nonsensical sub-tokens. While Latin BERT improved upon this, it still lagged behind the CRF.
- Computational Efficiency: The CRF model was trained in just 21 minutes on a standard desktop CPU. In contrast, the LLMs required a dedicated server with four NVIDIA A40 GPUs. This demonstrates that high-performance results in the digital humanities do not always require prohibitive hardware costs.
- Scale of Output: The final automated system provided a 9.1x multiplier on human effort, expanding the small expert-labeled set into a comprehensive index of 41,784 legal references across the entire five-book collection.
Discussion and Future Work
While highly successful, the model encountered specific challenges. It initially struggled with references to the Digest (abbreviated as "ff."), requiring a rule-based post-processing step. Additionally, the system occasionally flagged references to the Gospels as legal citations due to their similar syntactic structure. While technically false positives in a legal context, these findings offer interesting cross-disciplinary insights for religious studies.
The authors have made the resulting index publicly available. Future work will focus on refining erroneous annotations and applying this pipeline to other major corpora within the Corpus Iuris Canonici and Corpus Iuris Civilis, ultimately aiming to map the entire landscape of medieval legal authority.
Summary updated