Researchers test retrieval models for finding Church Father citations in Bullinger's letters
An arXiv paper reports the first computational approach to detecting patristic references in the Swiss Reformer's correspondence, combining lexical and dense retrieval with topic modelling.
In plain words
Heinrich Bullinger, a 16th-century Swiss Reformer, filled his letters with citations of the Church Fathers, and historians had no automated way to find them. The authors tested search methods that match citation text against source works, plus topic models that group letters by theme. BM25, a keyword-based search method, worked best for explicit citations, while fusing it with neural retrieval helped find subtler, implicit references.
What was announced
The authors compared lexical BM25, dense bi-encoders, and hybrid methods on both explicit and implicit references. BM25 led on explicit citations with 94% Recall@20, while a Latin encoder fine-tuned on biblical reuse was the strongest bi-encoder, reaching 82% on explicit and 16% on implicit references. Because the two methods recovered largely different implicit references, fusing their rankings lifted implicit Recall@20 to 22%, close to cross-encoder reranking at 26% but without its inference cost.
For topic modelling, the team extended FASTopic with a seed-guided loss function and editorial abstract embeddings. Coherence gains over the baseline were minimal, and dominant topic weights barely exceeded a uniform distribution, so topics could not inform retrieval. Still, qualitative and temporal analysis showed some topics tracking documented historical events, and a pilot of LLM-based concept induction (LLooM) suggested a more human-interpretable alternative. The authors present this as a first approach to the task.
What it does not show
- The evidence is the paper's abstract only; full experimental details, datasets, and evaluation setup are not available for verification.
- Reported metrics come from the authors' own evaluation and have not been independently confirmed.
- The authors state topic weights were too weak to inform retrieval, so the topic modelling component has unclear practical value.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.