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

TL;DR

The study evaluates computational methods for detecting biblical intertextual references in Karen Blixen's Seven Gothic Tales using a manually annotated benchmark of 189 references. It compares traditional retrieval models like BM25 and TF-IDF with dense semantic models, including Danish sentence encoders fine-tuned with hard negatives and cross-validation. Fine-tuning significantly improves performance, especially on allusions, and models serve as heuristic co-readers rather than definitive discovery tools.

Why It Matters

This work advances the use of AI in literary analysis by demonstrating how retrieval models can support scholarly interpretation of intertextuality, particularly in complex, transformed textual references. It highlights the importance of human-in-the-loop validation, showing that model outputs can suggest meaningful connections beyond existing annotations.

Key Findings

• Linguistically normalized BM25 achieves strong zero-shot performance with an R@10 of 0.365 and retrieves all direct quotations within the top ten results.
• The best zero-shot dense model matches BM25 in overall R@10 (0.360) but outperforms it on allusions, indicating better handling of indirect references.
• Fine-tuning the DFM-large Danish encoder with hard negatives and cross-validation increases R@10 from 0.265 to 0.508 and improves allusion retrieval from 0.138 to 0.339, while expert evaluation found meaningful references among predicted false positives.