Retrieving Biblical Intertextual References in Karen Blixen's Seven Gothic Tales

What it claims

Lexical retrieval is still a strong baseline for finding biblical references in historical literary Danish. Linguistically normalized BM25 reaches Recall@10 of 0.365 over 189 annotated references. It places every quotation in its top ten. The best zero-shot dense encoder scores 0.360 overall but does better on allusions, at 0.165 R@10. Fine-tuning DFM-large with hard negatives raises it from 0.265 to 0.508 overall, and on allusions from 0.138 to 0.339. That is roughly double the best zero-shot model on the hardest category. The practical point: sparse retrieval is enough for near-verbatim reuse, and a few hundred labelled pairs of task-specific training are what move the needle on transformed references.

How it works

• Queries. Passages from the 1935 Danish first edition of Syv fantastiske Fortællinger, in original orthography. The median length is 7 words, ranging from one word ("Morderengel") to 204 words.
• Labels. Taken from a peer-reviewed critical-edition commentary compiled by a single editor. 23 annotations that cite two or three verses were split into separate pairs, giving 189 pairs.
• Corpus. The 1871 Old Testament plus the 1907 New Testament, 31,170 verses in total. These were chosen as historically plausible for Blixen's period. The edition itself quotes the 1992 translation. The authors argue modern translations erase lexical cues, for example "Hil dig, jødekonge" becomes "Længe leve jødernes konge".
• Methods compared.
• TF-IDF and BM25.
• Multilingual and Danish sentence encoders.
• Each tested on both raw and linguistically normalized text.
• Fine-tuning. Only the strongest Danish encoder (DFM-large) is fine-tuned, using hard negatives and five-fold cross-validation.
• Strata. Results are broken down into quotation, paraphrase and allusion. These categories are assigned automatically from lexical overlap, not by hand.

Evidence

• Normalized BM25: R@10 0.365, with all quotations retrieved in the top 10.
• Best zero-shot dense model: R@10 0.360 overall and 0.165 on allusions.
• DFM-large: 0.265 zero-shot and 0.508 after fine-tuning (five-fold CV). On allusions it goes from 0.138 to 0.339.
• Scholar review: a literary scholar checked 30 selected rank-one predictions that the benchmark scored as false positives, and judged 7 of them to be meaningful additional references.

Limits

• Incomplete ground truth. The annotations come from one editor and document established references rather than every defensible one. Measured recall and precision are therefore understated, as the 7 of 30 result shows.
• Unknown source text. Which Bible Blixen actually used is unknown. She may have worked from Danish or English, since the tales were first written in English, or from cultural memory. The chosen corpus is plausible, not confirmed.
• Duplicate wording within the Bible. The Bible repeats itself, so the same motif can appear in several verses or in both testaments. A retriever unaware of theological citation conventions may surface an Old Testament source where an editor would cite the New Testament, or the reverse.
• Proxy categories. The quotation, paraphrase and allusion strata are automatic proxies based on lexical overlap.
• Small benchmark. It has 189 pairs from a single work. Fine-tuning gains are measured by cross-validation within it.
• No release details. The fetched text is truncated partway through Section 4. It states no code, weights, data release or compute figures. The paper itself is licensed CC BY 4.0.
• Intended use. The authors frame the output as candidates for expert close reading, not as discoveries.