Source-linked AI summary
Pair-Level Essay-Scale Republication and Reuse from Fragmented Historical Text Reuse: A Workflow Study on Eighteenth-Century Books and Newspapers
Ke Shu, Kira Hinderks, Eetu Mäkelä, Mikko Tolonen
TL;DR
The paper tackles recovery of essay-scale republication and reuse from fragmented text-reuse hits, where the key challenge is consolidating pair-level evidence into transmission relations. It evaluates a staged rule-based workflow against classifier, LLM, and automated adaptation approaches, producing compact auditable candidate spaces and confirming all 176 newspaper predictions as genuine.
Problem
The study addresses how to recover essay-scale republication and reuse across eighteenth-century books and newspapers when fragmented matches must be consolidated into transmission relations.
Method
The paper compares a staged rule-based workflow with decision-tree, direct-LLM, and automated rule-adaptation approaches over shared pair-level evidence.
Results
The staged workflow provides the strongest overall labeled ECCO–ECCO precision-recall balance and yields a compact candidate space whose 176 newspaper predictions were manually confirmed as genuine.
Takeaways & Limitations
Under incomplete positive coverage, auditable pair-level evidence consolidation offers a practical basis for compact candidate spaces suitable for historical inspection.
Takeaways & Limitations
Because the corpus centers on Hume, the findings have limited direct generalizability to other authors and historical periods.
Abstract
from arXiv · showhide
This paper addresses the recovery of essay-scale republication and reuse from fragmented text-reuse evidence, a setting whose central challenge is pair-level evidence consolidation and not fragment retrieval alone. The study focuses on a candidate set centered on essays by eighteenth-century Scottish philosopher David Hume, spanning books from ECCO (Eighteenth Century Collections Online) and historical newspapers. Because the input consists of fragmented reuse hits instead of clean document pairs, and positive coverage is inherently incomplete, we formulate the task as pair-level evidence consolidation into plausible transmission relations and compare three methodological families: a staged rule-based workflow, baselines (a decision tree and two direct LLM settings), and automated rule adaptation. On labeled ECCO--ECCO slices, pair-level feature aggregation alone already reaches 0.948 F1 on the main labeled slice, while the final workflow gives the strongest overall precision-recall trade-off among the tested rule stages. On the full ECCO--ECCO candidate universe, direct LLM baselines flag up to 14,886 pairs as reprints compared to 771 for the final workflow, behaving in this direct-prompt setup as high-recall candidate expanders rather than precision-controlled deployment classifiers. On ECCO--Newspaper, manual audit confirms all 176 predicted positives as genuine cases of republication or reuse, while issue duplication and source-side multiplicity reveal additional provenance structure. Under incomplete ground truth, auditable pair-level evidence consolidation provides a practical way to produce compact candidate spaces for historical inspection.
1 Introduction
The paper frames essay-scale republication detection as pair-level consolidation of fragmented reuse evidence across eighteenth-century books and newspapers, rather than fragment retrieval alone. It studies a workflow over Hume’s essays, ECCO books, and historical newspapers, comparing staged rules, baselines, and automated rule adaptation.
- 1 Introduction: The study defines essay-scale reuse around substantial preservation of Hume source essays despite paragraph reordering, abridgment, rewriting, and cross-media segmentation.The operational definition is designed to capture genuine reuse beyond short overlaps or perfectly identical copies.
- 1 Introduction: The central problem is consolidating noisy, high-recall fragment matches into assessable transmission relations, because extensive overlap does not automatically establish reprint.The task is framed as evidence consolidation rather than supervised classification or fragment retrieval.
- 1 Introduction: The workflow study uses 17 Hume books from ECCO and compares staged rules, a decision tree, direct LLM settings, and automated rule adaptation across ECCO books and Burney Newspapers.All methodological families operate over the same pair-level evidence space, representing increasing degrees of automation.
- 1 Introduction: The contribution extends fragment-level retrieval systems such as Passim toward pair-level transmission judgment under incomplete positive coverage, while connecting ECCO books with eighteenth-century newspapers.Related systems including Viral Texts and Reception Reader [11] primarily emphasize retrieval and exploration rather than pair-level evidence consolidation.
2 Task, Data, and Methods
The study consolidates fragmented BLAST-style reuse hits into pair-level evidence for detecting essay-scale republication and reuse. It compares an auditable staged rule workflow with decision-tree, direct-LLM, and automated-adaptation approaches across ECCO books and newspapers.
- Task and Features: The task aggregates fragment-level hits into document-pair evidence using coverage, span, section distribution, chaining, and title, heading, quotation, and paratext cues.Several nearby fragments can jointly support one reused passage even when no individual hit is decisive.
- Task and Data: The dataset centers on 17 books by Scottish philosopher David Hume and uses fragment signals to reduce millions of ECCO pair possibilities to tens of thousands.This contraction makes expert curation feasible and makes candidate-space control central to discovery quality.
- Rule-based workflow: The retained workflow is a fixed, inspectable four-stage rule cascade progressing from shallow span heuristics through structural and context-aware rules to hard-case rescue and suppression.Each positive decision is traceable to a structural, coverage, context, or rescue gate; the rescue branch narrowly admits low-fanout near misses without quotation or paratext cues.
- Baselines: Baselines test the same evidence space with a decision tree trained once on 60 balanced discovery pairs and two direct Qwen3-30B-A3B-Instruct-2507 [9] settings.The LLM variants are text-only and structured; they are treated as candidate-expansion reference conditions because they may use pretrained knowledge beyond fixed features.
- Automated adaptation: Automated adaptation proposes bounded rescue or suppression rules from stage-to-stage hard-case differences, providing a reproducible middle ground between broad LLM prompting and labor-intensive hand tuning.When used, the LLM proposes candidate rules rather than serving as the final classifier.
- Evaluation: Evaluation covers labeled ECCO–ECCO slices, deployment behavior on the full ECCO–ECCO universe, and ECCO–Newspaper cases, with the hard slice treated as a boundary diagnostic rather than fully held-out generalization.The final workflow’s rescue and suppression rules were refined using hard-case pools overlapping that split.
3 Results
Results show that pair-level evidence aggregation performs strongly on labeled ECCO–ECCO data, while the final workflow offers the best overall balance and controlled deployment output. In ECCO–Newspaper, manual audit confirmed every predicted positive as genuine republication or reuse despite incomplete ground truth.
- ECCO–ECCO labeled slices: 0.948 F1 for structural-only aggregation on the main labeled slice shows that pair-level feature aggregation is already strong, while the final workflow provides the strongest overall balance.On hard cases, decision-tree F1 fell to 0.063, while naive and structural-only workflows remained at 0.171.
- Rule-based progression: 0.263 hard-slice F1 after adding contextual cues, versus 0.171 before them, shows that context supplies the main rule-based gain beyond structural aggregation.The final stage primarily adds boundary control, raising precision from 0.312 to 0.333 and reducing deployment output from 1,265 to 771 positives.
- Method-family comparison: 0.455 hard-slice recall and 0.317 F1 make automated adaptation more expansive than the final workflow, which records 0.227 recall and 0.270 F1.The decision tree is informative overall but remains brittle on difficult cases.
- Full candidate-universe deployment: 771 final-workflow positives versus up to 14,886 for direct LLM baselines show the final method’s most controlled output on the full ECCO–ECCO candidate universe.Direct LLM baselines produced the broadest outputs and highest recall on every labeled split, while the final workflow maintained controlled output.
- ECCO–Newspaper audit: 176 of 176 newspaper cases predicted as republication or reuse were manually confirmed genuine, while a diagnostic audit of 49 predicted negatives found 8 reprints and 38 non-reprints.Because the negative audit oversampled difficult cases, its false-negative rate is not a population-prevalence estimate; the setting required deployment-and-audit evaluation without pre-existing ground truth.
4 Analysis
The analysis finds that pair-level aggregation dominates performance, while staged rule refinement mainly improves boundary control on ambiguous pairs. Domain-specific transmission patterns and auditability qualify the results, which remain limited by Hume-centered data and annotation constraints.
- 4 Analysis: The final workflow achieves the highest all-labeled F1 and hard-case precision, while automated adaptation sacrifices precision for markedly higher hard-case recall.A shallow decision tree reaches 0.909 F1 on the main labeled slice but falls to 0.063 on hard cases; direct LLM inference produces broader positive outputs.
- 4 Analysis: ECCO–ECCO review confirmed 81 of 104 apparent false positives as genuine cases outside current ground truth, showing that strict full-corpus precision is conservative.The confirmation rate is 78%, or 86% when uncertain cases are excluded.
- 4 Analysis: The workflow remains auditable because coverage ratio, bundle structure, and chain structure expose complementary evidence for each prediction.These signals indicate source coverage, target concentration, and fragment coherence independently of the classification boundary.
- 4 Analysis: Newspapers require historically specific interpretation because reused essays may appear as embedded extracts, recomposed digests, or verbatim reprints attributed to another source.These forms make essay-scale textual identity a historical transmission question rather than only a data-quality problem.
- Limitations: Generalizability is limited because the corpus centers on Hume, while non-blind newspaper annotation, absent inter-annotator agreement, OCR variation, and direct-prompt LLM conditions constrain evaluation.The study is therefore a historically motivated stress test rather than an author-independent benchmark.
5 Conclusion
The study presents a workflow for essay-scale republication and reuse detection across ECCO books and eighteenth-century newspapers, using pair-level evidence consolidation to reduce candidates while preserving useful recall.
- Conclusion: The workflow addresses essay-scale republication and reuse detection across two structurally distinct historical environments: ECCO books and eighteenth-century newspapers.ECCO supports quantitative evaluation over partially labeled pairs, whereas newspapers require deployment-and-audit without prior ground truth.
- Conclusion: Staged pair-level evidence consolidation produced a compact, auditable candidate space while preserving useful recall.
- Conclusion: ECCO books enable quantitative evaluation over partially labeled pairs, while eighteenth-century newspapers require deployment-and-audit in the absence of prior ground truth.
GenAI Disclosure
The authors used ChatGPT and Claude for writing and editing assistance, while retaining sole responsibility for the paper’s scientific content and conclusions.
- GenAI Disclosure: ChatGPT and Claude assisted with writing and editing, but the authors alone were responsible for the scientific content, experimental design, results, and conclusions.