PaperlayerView exact revision on arXiv

Source-linked AI summary

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

András Kovács, Alexander Conroy, Daniel Hershcovich, Jens Bjerring-Hansen

arXiv:2609.35765v1cs.CL

TL;DR

Literary intertexts can be difficult to retrieve when wording is transformed through paraphrase or allusion and separated by historical language and translation. This study benchmarks sparse and dense retrieval for biblical references in Karen Blixen’s Seven Gothic Tales, then fine-tunes a Danish encoder. Fine-tuning improves overall retrieval, while editorial annotations remain selective and model suggestions are candidates for expert evaluation.

  • Problem

    Retrieving biblical intertexts is difficult when Blixen’s antiquated Danish, transformed allusions, historical translations, and overlapping biblical sources obscure textual matches.

    Source: §1 · pp. 1–2

  • Method

    The study builds a 189-reference benchmark from scholarly commentary, compares sparse and dense retrieval over 31,170 historically plausible Danish Bible verses, and fine-tunes a Danish sentence encoder.

    Source: §3 · p. 4, §9 · p. 15

  • Results

    Fine-tuning dfm-large raises overall R@10 from 0.265 to 0.508, with the strongest gains for paraphrases and allusions.

    Source: §9 · pp. 15, 17

  • Takeaways & Limitations

    Retrieval models can recover documented references and propose interpretive candidates for scholarly evaluation, rather than making automatic discoveries.

    Source: §9 · p. 17

    The editorial benchmark is selective rather than an exhaustive inventory of plausible intertexts, so unannotated retrieved verses may still be relevant.

    Source: §3 · p. 4, §8.2 · p. 14

Abstract

from arXiv · show

Identifying intertextual references is central to literary scholarship, but computationally difficult when source material is transformed through paraphrase, allusion, historical language, and translation. We investigate this problem through biblical intertextuality in Karen Blixen's Seven Gothic Tales. Drawing on the commentary to a critical edition, we construct a benchmark of 189 annotated references and evaluate retrieval against all 31,170 verses of historically plausible Danish Old and New Testament translations. We compare TF-IDF and BM25 with multilingual and Danish sentence encoders, examine the effect of linguistic normalization, and fine-tune a Danish encoder using hard negatives and five-fold cross-validation. We analyze performance across automatically derived lexical-overlap strata representing quotations, paraphrases, and allusions. Linguistically normalized BM25 provides a strong zero-shot baseline, attaining an overall R@10 of 0.365 and retrieving every quotation within its ten highest-ranked verses. The best zero-shot dense model achieves a comparable overall score of 0.360 while performing better on allusions. Fine-tuning DFM-large raises its overall R@10 from 0.265 to 0.508 and more than doubles its performance on allusions, from 0.138 to 0.339. However, evaluation against editorial annotations alone understates the model's scholarly usefulness: a literary scholar judged seven of 30 selected rank-one predictions counted as false positives to be meaningful additional references. These findings show both the potential and the epistemic limits of computational intertextual retrieval. Rather than treating scholarly annotations as exhaustive or model outputs as discoveries, we propose retrieval models as heuristic co-readers that recover documented references and generate candidates for expert-led close reading.

1 Introduction

Biblical intertexts in Seven Gothic Tales range from quotations to transformed allusions, complicated by historical Danish and competing source passages. The study benchmarks retrieval against editorial references and finds that fine-tuning improves retrieval, while model suggestions can also support expert-led interpretation.

1 Introduction

  • Biblical references range from recognizable quotations to transformed allusions, complicated by antiquated Danish, long sentences, historical translations, and overlapping biblical sources.Blixen’s wording may differ substantially from modern translations, while similar phrases, events, or motifs may occur in multiple biblical locations.Source: §1 · pp. 1–2
  • The 189-reference benchmark uses scholarly commentary as an expert basis, but its annotations are not an exhaustive inventory of defensible connections.The edition documents established references, so a retrieved verse outside its annotations may still be relevant to literary interpretation.Source: §1 · p. 2
  • The study compares sparse and dense retrieval on 31,170 verses from historically plausible Danish translations, with normalization, fine-tuning, and analysis by reference type.It uses 1871 Old Testament and 1907 New Testament translations and separates quotations, paraphrases, and allusions.Source: §1 · p. 2
  • Fine-tuning raises overall Recall@10 to 0.508, while normalized BM25 scores 0.365 and retrieves every quotation in its top ten verses.The best zero-shot dense model scores 0.360; fine-tuning raises allusion Recall@10 from 0.165 to 0.339.Source: §1 · p. 2
  • Retrieval is framed as a heuristic co-reader that recovers documented references and proposes alternatives for scholarly examination, not as automated editorial annotation.Some apparent false positives may be plausible alternative references, including links to earlier Old Testament sources.Source: §1 · p. 2

2 Related Work

Prior work spans interpretable lexical matching, semantic representations, contextual fine-tuning, and generative interpretation. This study addresses limited systematic comparison across reference types and historically variable literary and source languages.

2 Related Work

  • Lexical matching offers interpretable evidence and works well for quotations, but observable-overlap reliance makes paraphrases and allusions harder to detect.Tesserae uses shared words and related lexical features; TRACER supports configurable similarity measures, and Duan [6] surveys hybrid approaches.Source: §2 · p. 3
  • Lemmatized Latin Word2Vec improved synonym detection and intertextual search, but static embeddings lack context and require word similarities to be aggregated at passage level.Burns et al. illustrate how distributional representations can recover relations beyond exact overlap.Source: §2 · p. 3
  • Contextual fine-tuning with hard negatives can distinguish semantic equivalence from superficial overlap, including when manually annotated data are scarce.D’Angelo et al. applied contrastive learning to semantic reuse in Ancient Greek.Source: §2 · p. 3
  • Generative models have produced novel intertextual observations, but false dependencies, long queries, and over-interpretation leave expert evaluation important.Expert-informed prompts can improve theoretical grounding while also increasing over-interpretation.Source: §2 · p. 3
  • Prior work rarely compares lexical and dense retrieval on one expert-annotated literary benchmark spanning quotations, paraphrases, allusions, and historically variable languages.The study compares these methods for biblical references in Karen Blixen’s Syv fantastiske Fortællinger, including preprocessing and task-specific fine-tuning.Source: §2 · p. 3

3 Source Texts and Benchmark Construction

The benchmark draws on a scholarly Danish edition and constructs independently evaluated Blixen–Bible pairs. Its source translations are selected for historical plausibility, while the editorial annotations remain selective.

3 Source Texts and Benchmark Construction

  • The scholarly edition preserves original orthography and punctuation, exposing models to Blixen’s antiquated style and long hypotactic sentences.The edition is based on the 1935 Danish edition and includes limited emendations supported primarily by the manuscript and Blixen’s own corrected copy.Source: §3 · p. 4
  • The peer-reviewed commentary is the most thorough documented search to date, but a single editor’s work may have overlooked references and is not definitive.The apparatus provides the benchmark without claiming to account for every possible biblical reference.Source: §3 · p. 4
  • The benchmark splits multi-verse annotations into independent references, yielding 189 Blixen–Bible reference pairs.The editor identifies 23 cases referring to two or three distinct biblical verses; splitting avoids making them inherently more difficult than single-target cases.Source: §3 · p. 4
  • Annotated Blixen passages range from one word to 204 words, creating substantial variation in query length.The median passage length is seven words.Source: §3 · p. 4
  • The study uses the 1871 Old Testament and 1907 New Testament as a historically plausible, consistent corpus of 31,170 verses, not as definitive evidence of Blixen’s sources.Blixen’s preferred Bible translation and whether her references arose from direct reading are unknown; the edition also points to several older translations.Source: §3 · pp. 4–5
  • Historical wording can preserve intertextual cues that disappear in newer translations, as illustrated by the changed Danish rendering of Jesus’s mock salutation.The proposed 2036 translation changes “Hil dig, jødekonge” to “Længe leve jødernes konge,” removing the lexical cue to Grundtvig’s hymn.Source: §3 · p. 5

4 Operationalizing Biblical Intertextuality

The study adapts theories of textual relationships into a three-part taxonomy and evaluates a deliberately narrower retrieval task. Classification remains interpretive, and benchmark disagreement does not establish that an intertext is absent.

4 Operationalizing Biblical Intertextuality

  • The taxonomy adapts transtextuality and manifest-intertextuality frameworks rather than directly adopting them as a computational annotation scheme.Genette offers a taxonomy of textual relations; Fairclough distinguishes explicit incorporation from broader conventions and discourses.Source: §4 · p. 5
  • The taxonomy distinguishes quotations that reproduce distinctive wording, paraphrases that reformulate a localized source, and allusions that evoke a passage, event, figure, or motif indirectly.Broader work-level strategies such as parody and pastiche are excluded, though localized overlaps within them may fit the categories.Source: §4 · pp. 5–6
  • Boundaries between paraphrase and allusion are blurred, and identifying intertexts depends on readerly literary competence that remains challenging to replicate in NLP.Riffaterre describes recognition as resolving semantic anomalies through an underlying source structure.Source: §4 · p. 6
  • The task must accommodate lexical, semantic, narrative, and culturally mediated relations across passages of markedly different lengths and linguistic forms.Antiquated style and historical and translational distance can obscure similarity between Blixen’s text and its biblical source.Source: §4 · p. 6
  • The retrieval task ranks Danish Bible verses for an annotated Blixen passage and tests recovery of the commentary’s source, not intertextuality in general.Given the commentary’s selectivity, missing the annotated verse signals benchmark disagreement rather than absence of a plausible relationship.Source: §4 · p. 6

5 Reference Distribution and Automatic Classification

The editorial commentary’s 161 biblical annotations were split where needed into 189 reference instances. The section describes their distribution and classifies them into lexical-overlap strata for quotations, paraphrases, and allusions.

5.1 Reference Counts and Distribution

  • 23 multi-target annotations were split into 28 additional instances, bringing the benchmark to 189 pairings of a Blixen lemma and a continuous biblical passage.The 161 annotations came from 164 Bible-related annotations after three apocryphal references were excluded.Source: §5.1 · p. 6
  • 106 instances target one verse and 83 target passages of two to ten verses; Blixen lemmas have a median of 7 words versus 34 for biblical passages.Biblical passages have median lengths of 26 words for single-verse targets and 57 for multi-verse targets.Source: §5.1 · p. 7
  • Syndfloden over Norderney contributes 63 instances; Matthew has 71 and Genesis 26, while 24 of 66 biblical books are represented.The 189 flows are grouped by testament in Figure 1; 65 instances target the Old Testament and 124 the New Testament.Source: §5.1 · p. 7, Figure 1 · p. 8

5.2 Automatic Classification of Relationship Types

  • The classification pipeline lemmatizes and stems both texts, then uses Jaccard overlap and thresholds of 0.1 and 0.3 to assign 109 allusions, 54 paraphrases, and 26 quotations.The score ranges from 0 for no shared processed tokens to 1 for identical token sets.Source: §5.2 · p. 7
  • The labels are reproducible lexical-overlap approximations, not expert-validated literary classifications; Jaccard can penalize longer passages and cannot establish relationship type.Lemmatization and stemming may also collapse meaningful linguistic distinctions.Source: §5.2 · p. 9

6 Experimental Setup

The experiments rank all 31,170 Bible verses for each annotated Blixen query, comparing sparse and dense zero-shot retrieval with task-specific fine-tuning. Preprocessing, five-fold evaluation, and several ranking metrics define how those comparisons are made.

6 Experimental Setup

  • Each of 189 Blixen queries ranked all 31,170 verses, with the continuous editorially annotated passage as the target.For multi-verse targets, any verse in the passage counted as relevant, and retrieval succeeded if at least one appeared in the results.Source: §6 · p. 9

6.1 Preprocessing

  • Orthographic normalization converts historical aa to modern å; sparse retrieval also compares surface text with lemmatized-and-stemmed text, while dense models use normalized surface text.At least one aa-to-å conversion occurred in 57% of Bible verses.Source: §6.1 · p. 9

6.2 Sparse Retrieval

  • TF–IDF and BM25 were evaluated on both surface-form and linguistically normalized representations; BM25 also normalizes for document length.BM25 applies term-frequency saturation to reduce preference for longer verses based merely on more query-term matches.Source: §6.2 · p. 9

6.3 Dense Retrieval

  • Zero-shot dense retrieval compares Multilingual E5 and Danish Foundation Models encoders by cosine similarity between independently encoded queries and verses.The documentation does not identify pretraining data at the level of individual works, so zero-shot results do not guarantee absence of pretraining overlap.Source: §6.3 · p. 10

6.4 Fine-Tuning

  • DFM-large was fine-tuned with multiple-negatives ranking loss and mined hard negatives, using five annotation-grouped folds and evaluating each held-out query against the full corpus.The folds assess generalization to unseen annotated Blixen passages, not necessarily to unseen Bible verses, since verses may recur across annotations.Source: §6.4 · p. 10

6.5 Evaluation Metrics

  • The study macro-averages P@1, R@10, MRR@10, and nDCG@10 over the 189 instances, measuring top-rank hits, top-ten hits, reciprocal rank, and rank-sensitive relevance.R@10 is an instance-level hit measure, not conventional set recall.Source: §6.5 · p. 11
  • For multi-verse targets, R@10 records whether any target verse appears in the top ten, whereas nDCG@10 also reflects how many relevant verses are retrieved and their ranks.Each retrieved verse is binary relevant if it belongs to the query’s gold-standard set.Source: §6.5 · p. 11

7 Results

The results compare sparse and dense zero-shot retrieval, test fine-tuning, and reassess selected apparent errors; fine-tuning yields the strongest overall scores, while expert reassessment qualifies annotation-based error counts.

7.1.1 Sparse Retrieval

  • 0.365 R@10: normalized BM25 was the strongest sparse setup, retrieved all 26 quotations in its top ten, and outperformed other sparse configurations on paraphrases and allusions.Normalization improved BM25 overall R@10 from 0.286 to 0.365, but TF–IDF paraphrase R@10 fell from 0.500 to 0.426.Source: §7.1.1 · p. 11, Table 1 · p. 12

7.1.2 Dense Retrieval

  • 0.360 R@10: e5-large led zero-shot dense retrieval and allusions (0.165), but remained just below normalized BM25’s 0.365 overall.Large variants beat their small counterparts within both model families, but architecture and training data also differed.Source: Table 2 · p. 12, §7.1.2 · p. 12

7.2 Effect of Fine-Tuning

  • 0.508 R@10: fine-tuned dfm-large rose from 0.265 and led every metric, with its largest gains on paraphrases (0.315 to 0.685) and allusions (0.138 to 0.339).BM25 remained stronger on quotations for R@10, MRR@10, and nDCG@10; the systems tied at quotation P@1 of 0.654.Source: §7.2 · p. 12, Table 3 · p. 13

7.3 Expert Reassessment of Apparent False Positives

  • 7 of 30 selected dfm-ft rank-one errors were judged relevant (P@1 0.233), including 2 of 7 paraphrase cases and 5 of 23 allusion cases.The cases were selected from apparent errors rather than sampled independently, so the findings do not estimate full-benchmark performance or support comparisons with other retrieval methods.Source: §7.3 · pp. 13–14, Table 4 · p. 13

8 Analysis and Discussion

Fine-tuning improved retrieval across heterogeneous references, but editorial annotations are selective, and biblical self-reference makes the meaning of a match depend on the study’s interpretive aim. In this dataset, retrieval success showed no clear association with reference length.

8.1 Effects of Fine-Tuning

  • At Job 40:10–19, dfm-large ranked the annotated passage’s first verse 430th, while dfm-ft ranked a verse from the passage first, illustrating fine-tuning’s gains on paraphrases and allusions.The example does not establish that fine-tuning resolved length discrepancies or reveal which textual features drove the improvement.Source: §8.1 · p. 14

8.2 The Selectivity of the Ground Truth

  • Seven of 30 selected dfm-ft predictions counted as P@1 errors were judged relevant, showing that benchmark errors can include unrecorded references.The examples were selected from apparent errors, so they do not provide an adjusted performance estimate.Source: §8.2 · p. 14

8.3 Biblical Self-Reference and Interpretive Conventions

  • For “the new Jerusalem,” dfm-ft retrieved Isaiah rather than the annotated Revelation passage, a match that may still matter for the formulation’s textual history.The model appears to prioritize lexical and semantic similarity without representing the convention that an Old Testament prophecy may be cited through its New Testament fulfilment.Source: §8.3 · p. 15
  • Whether an alternative biblical parallel counts as useful depends on whether the study seeks editorial source attributions or possible intertexts; scholars must assess its literary significance.Retrieval criteria should reflect the study’s interpretive objective rather than universally enforcing or disregarding theological conventions.Source: §8.3 · p. 15

8.4 Reference Length and Retrieval Performance

  • Across this dataset, retrieval success showed no clear relationship to lemma length, biblical-passage length, or whether the two texts had similar lengths.The ten-verse Job example does not establish that fine-tuning generally favours long passages or eliminates length-related difficulties.Source: §8.4 · p. 15

9 Conclusion and Future Work

Fine-tuning a Danish sentence encoder improved biblical-intertext retrieval, especially for paraphrases and allusions, while expert review showed that some benchmark errors were meaningful additional references. The study presents retrieval as a heuristic aid for scholarly close reading, with future work directed toward stronger training and evaluation.

9 Conclusion and Future Work

  • Fine-tuning raised dfm-large’s overall R@10 from 0.265 to 0.508, with the strongest improvements for paraphrases and allusions where direct lexical overlap is limited.The study reports that this produced its strongest overall results across the reported metrics.Source: §9 · pp. 15, 17
  • Figure 2 shows no clear relationship between reference lengths and whether dfm-ft retrieved a ground-truth verse within its top ten.Each point represents one reference; shape indicates relationship type and colour indicates retrieval success.Source: Figure 2 · p. 16
  • Seven of 30 selected dfm-ft predictions counted as errors were judged meaningful additional references, demonstrating that editorial annotations are not an exhaustive inventory of relevant relationships.Such model suggestions are interpretive candidates for scholarly evaluation, not automatic discoveries.Source: §9 · p. 17
  • Future work should test additional fine-tuning approaches and use a larger expert-annotated dataset to compare methods more reliably and evaluate relationship types against literary judgements.The study also calls for systematic investigation of possible retriever bias toward either testament.Source: §9 · p. 17
  • A retrieval-augmented system could propose passages with evidence-grounded explanations for scholars to inspect, reject, or refine, leaving final interpretive authority with the reader.The paper frames intertextual retrieval as a demanding benchmark for language understanding beyond standard semantic-similarity tasks.Source: §9 · p. 17
Loading 2609.35765v1…