GVD: Governed Versioning and Deduplication for Document Repositories
Mohammadreza Sediqin, Mohammadreza Sediqin, Shivali Dalmia, Shivali Dalmia, Sumukha Thoppanahalli, Sumukha Thoppanahalli, Abhishek Mukherji, Abhishek Mukherji
cs.AI
Quick verdict
GVD unifies version-family linking and rule-level conflict governance without LLMs; strong F1 on a private enterprise corpus but not independently reproducible.
Summary
Document repositories accumulate revised, duplicated, and contradictory guidance, but prior work treats versioning, deduplication, and contradiction detection as isolated pairwise tasks. GVD unifies them into a governed ingestion pipeline built entirely from local, encoder-scale models: all-MiniLM-L6-v2 embeddings and a DeBERTa-v3-large NLI model. Stage 1 assigns incoming documents to version families via centroid retrieval plus bidirectional rule alignment (dual directional coverage above threshold). Stage 2 compares each new rule against the family memory: a cosine pre-filter, bidirectional NLI labeling pairs as duplicate, contradiction, asymmetric refinement, or neutral, and training-free Counterfactual Span Probing (CSP) that reclassifies residual neutral pairs using span categories (value changes, polarity reversals, one-sided clauses). Relation-specific policies suppress duplicates, insert new knowledge, and route consequential changes to human review while retaining version lineage. On 120 confidential enterprise documents (~3,900 rules) replayed as 140 ingestions across 59 families, GVD reports F1 0.971 for family assignment (vs RETSim 0.943) and 0.94 for rule consistency (0.90 without CSP), with 381 of 4,396 decisions escalated to review. Main limitations: the corpus is proprietary and no code is released; CSP's span-extraction algorithm is underspecified; evaluation uses small balanced sets (300/500 pairs) without significance testing; and the routing policy is validated only qualitatively.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Bidirectional rule-coverage alignment assigns documents to version families more reliably than direct similarity scoring, reaching F1 0.971 versus 0.943 for RETSim (Table 4).
- Training-free Counterfactual Span Probing recovers duplicates and asymmetric refinements collapsed into the neutral class, raising rule-consistency F1 from 0.90 to 0.94 (Tables 2 and 5).
- Relation-specific update policies concentrate human effort on consequential changes, escalating 381 of 4,396 decisions while retaining an auditable version lineage (Secs 3.3 and 4.5-4.6).
- GVD is, to the authors' knowledge, the first framework to convert consistency relations into repository update policy over an evolving collection rather than stopping at pair classification (Sec 1).
Strengths
Problem framing goes beyond pairwise classification
Section 1 and 3.3 turn consistency labels into repository update policy (suppress, insert, quarantine, escalate) with version lineage, which the paper argues — plausibly — no prior versioning or NLI system does (Sec 4.7 discussion).
Component attribution is clean via staged ablations
Table 4 isolates matching design (centroid-only 0.837, one-directional 0.944, bidirectional 0.971) and Table 5 isolates Stage 2 (cosine 0.56, cosine+NLI 0.90, +CSP 0.94), so each claimed contribution has a measured effect.
CSP gains align with its stated mechanism
Baseline protocol is fairer than typical
Annotation quality and leakage control are reported
Weaknesses
No reproducible artifacts
CSP internals are underspecified
Evaluation sets are small, balanced, and partly circular
Routing-policy evaluation is anecdotal
Weak contextualization of the no-LLM design choice
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does GVD use a large language model (LLM)?
No. GVD runs entirely locally with two encoder-scale models: all-MiniLM-L6-v2 for rule embeddings and DeBERTa-v3-large fine-tuned on SNLI/MultiNLI for natural language inference. Its Counterfactual Span Probing step is training-free and rule-based, so no generative model is involved and document content never leaves the deployment environment.
What data was GVD evaluated on, and is the dataset public?
On 120 confidential enterprise guideline documents (PDF/DOCX/PPTX) from industrial clients under NDA, containing about 3,900 validated rules, replayed as 140 ingestions forming 59 version families. The dataset is not public, and the paper announces no code or data release, so results cannot be independently reproduced.
How accurate is GVD?
As reported: F1 of 0.971 for assigning documents to version families (beating RETSim at 0.943 and SBERT cosine at 0.704), and F1 of 0.94 for four-way rule-level consistency classification; the CSP step lifts this from 0.90 to 0.94, mostly on duplicate and neutral pairs. These numbers come from small expert-labeled sets (300 and 500 pairs) with no reported confidence intervals.
What problem does GVD solve that standard deduplication or NLI tools do not?
Standard tools label one pair of texts at a time and stop. GVD maintains a living repository: it links incoming documents into version families, compares each rule against the accumulated memory of the whole version lineage, then applies per-relation policies — suppressing duplicates, inserting new rules, and routing contradictions or value/polarity edits to human review — so every accepted change is traceable to the revision that introduced it.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (pipeline overview) is available only as a caption in the extracted text; the diagram itself could not be inspected.
- CSP's span-extraction internals are not described in the text, so judgments about that component rest on the authors' summary and Eq. 8 alone.
- All quantitative judgments rely on the paper's reported numbers; the corpus is private, so no external verification was possible.
- Table 1's reclassification examples appear garbled in extraction; their content was not used for substantive judgments.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.