GVD
A document corpus governance framework paper published on arXiv that proposes a unified approach to document versioning, deduplication, and conflict detection. It achieves auditable document collection governance through rule-level granularity analysis and counterfactual span probing techniques.
Timeline (last 90 days)
A paper titled "GVD: Governed Versioning and Deduplication for Document Repositories" was published on arXiv, proposing a framework that handles document version management, deduplication, and conflict detection in a unified manner.
GVD assigns new documents to version families via bidirectional rule alignment, and compares intra-document rules against family memory to identify four types of relationships: duplicates, contradictions, asymmetric refinements, and novel knowledge.
GVD determines the status of each individual rule relative to historical versions at the rule level, rather than at the level of the entire document.
GVD proposes Counterfactual Span Probing (CSP), a technique used to re-identify the true relationships between relevant rule pairs that have been misclassified as neutral by NLI models.
CSP improves the F1 score for rule consistency determination from 0.90 to 0.94.
GVD employs relationship-specific strategies: content identified as duplicate is automatically suppressed, and only changes with substantive impact are escalated for human review.
GVD fully preserves version lineage as an audit trail, enabling compliance attestation at the content level.
GVD's entire pipeline runs fully locally without using any large language models.
GVD's local approach is better suited to enterprise policy scenarios with standardized document structures and rule expressions, rather than cross-domain general-purpose knowledge base governance.
The evaluation in the GVD paper uses 120 enterprise documents processed as 140 ingestion runs, distributed across 59 version families.
2 more timeline events
All Facts (12)
GVD's evaluation scale (59 version families, 140 ingestion runs) is of medium magnitude, and its performance on larger-scale heterogeneous document repositories remains to be verified.
50%UnverifiedGVD assigns new documents to version families via bidirectional rule alignment, and compares intra-document rules against family memory to identify four types of relationships: duplicates, contradictions, asymmetric refinements, and novel knowledge.
50%UnverifiedGVD determines the status of each individual rule relative to historical versions at the rule level, rather than at the level of the entire document.
50%UnverifiedGVD proposes Counterfactual Span Probing (CSP), a technique used to re-identify the true relationships between relevant rule pairs that have been misclassified as neutral by NLI models.
50%UnverifiedCSP improves the F1 score for rule consistency determination from 0.90 to 0.94.
50%UnverifiedGVD employs relationship-specific strategies: content identified as duplicate is automatically suppressed, and only changes with substantive impact are escalated for human review.
50%UnverifiedGVD fully preserves version lineage as an audit trail, enabling compliance attestation at the content level.
50%UnverifiedGVD's entire pipeline runs fully locally without using any large language models.
50%UnverifiedGVD's local approach is better suited to enterprise policy scenarios with standardized document structures and rule expressions, rather than cross-domain general-purpose knowledge base governance.
50%UnverifiedThe evaluation in the GVD paper uses 120 enterprise documents processed as 140 ingestion runs, distributed across 59 version families.
50%UnverifiedGVD achieves an F1 score of 0.97 for version family construction and an F1 score of 0.94 for rule-level consistency.
50%UnverifiedA paper titled "GVD: Governed Versioning and Deduplication for Document Repositories" was published on arXiv, proposing a framework that handles document version management, deduplication, and conflict detection in a unified manner.
50%