A pipeline turns access rules into checked graphs
Text2NGAC: a multi-LLM pipeline for NGAC policy extraction with mutation-based verification · arXiv:2611.12572
In plain words
Access rules are often written in sentences, but computers need them in a formal structure. Text2NGAC is a language-model pipeline that untangles some unclear sentences, turns rules into a graph, and checks the graph’s structure.
Across four real policy collections, its extracted relationships matched the released annotations on nearly all scored sentences. That result applies to the collections’ original wording; a structural check can still miss a plausible error in meaning.
The research
Access rules written in ordinary language must be represented precisely before a system can enforce them.
Text2NGAC’s results give a concrete measure of how closely extracted relationships match existing annotations, while its verifier tests show which structural problems it catches and where a plausible meaning error can slip through.
Text2NGAC clarifies policy sentences, extracts all four NGAC relationship types into a graph, and checks the graph for structural problems.
The pipeline was evaluated on four real policy collections, with exact tuple match scored against released annotations for 714 sentences. On the collections’ original wording, accuracy ranged from 97.3% to 99.6% per collection.
An archived same-snapshot record also showed higher exact-match accuracy for the full pipeline than for single-model extraction across three collections.
The results support accurate extraction on these documents and show that specified graph corruptions can often be detected. They do not certify that a passing graph preserves the source policy’s intended meaning.
The extraction scores apply to the corpora’s original wording and released annotations. Structural checks can miss plausible meaning errors, and injected instructions in policy text can pass those checks.
What the evidence shows
- One pipeline, four rule types
Text2NGAC is a pipeline for extracting access rules from natural-language documents into NGAC, a graph-based access-control model. Existing approaches reviewed in the paper covered only parts of NGAC, and the paper identifies linguistic ambiguity and a lack of measured graph verification as further obstacles.
The pipeline addresses these together: it clarifies some sentences, extracts the model’s relationship types, and checks the resulting graph.
- Strong matches on original wording
On the collections’ original wording, the full pipeline matched released annotations closely.
In an archived same-snapshot production record, it outscored single-model R1 extraction, suggesting the added ensemble-plus-refinement stages helped in that record; the comparison does not isolate the ensemble from refinement.
- Structural checks have a boundary
The verifier detected most of the specified structural corruptions, but often missed parent-entity changes that remain plausible and well-typed. This contrast shows that strong detection of tested structural faults does not establish general semantic correctness.
- A targeted check and an enforcement replay
A separate same-graph comparison tested a provenance check for association edges with no supporting sentence. Detection rose from 0% with the check off to 100% with it on.
This result is specific to unsupported edges; it does not cover an incorrect edge that carries sentence provenance. A replay in the NIST reference implementation further showed that one extracted graph could be loaded and enforced, with its access decisions reproduced from the graph’s own NGAC meaning.
Deep
- Original paper authors
- Ayesha Habib, Huixiang Zhang, Dan Zhu, Waiz Haider
- Submitted to arXiv
- License
- arXiv non-exclusive