GraphCompliance: Structure-Aware Regulatory Compliance through Policy-Context Graph Alignment

arXiv:2510.26309 · cs.AI, cs.IR · Submitted 2025-10-30 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "GraphCompliance: Structure-Aware Regulatory Compliance through Policy-Context Graph Alignment".

Jane: The gist: GraphCompliance introduces a framework that represents regulatory texts as a Policy Graph and runtime contexts as a Context Graph,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: Alright, we’ve been talking about the core idea—this paper introduces GraphCompliance, which is all about making regulatory compliance more reliable by explicitly modeling the structure of the rules.

Jane: So, to put it simply, they are taking complex regulatory documents and turning them into a Policy Graph and treating runtime information as a Context Graph. They then align these two graphs using this Compliance Gate to anchor the reasoning of large language models in structured information.

Lu: The thesis is that by making the regulatory structure explicit through this policy graph, you can better handle cross-references and hierarchical dependencies that are usually messy in natural language documents.

Meng: So, the main claim they’re making is that this alignment process helps reduce the burden of interpretation for AI models because it gives them a clear roadmap.

Tom: That’s right. They are using deterministic structural analysis within the Compliance Gate to identify relevant regulatory requirements and evaluate them directly against the corresponding contextual evidence.

Jane: And they validated this by showing that their performance metrics, like MCC scores of zero point eight zero two for EU2UK and zero point seven nine three for CONTRACT, are numerically higher than those of the strongest compared methods on all three datasets.

Lu: Those results confirm that the proposed design helps preserve structural dependencies and organizes contextual facts in a way that is comparable across different regulatory settings, which is a key part of their contribution to this paper.

Tom: So, the essence is taking unstructured text and imposing structure onto both sides—the policy side and the context side—so the AI gets a constrained problem to solve.

Jane: And they’ve done quite a bit on that alignment process, showing that the key components of GraphCompliance contribute to its overall performance in their ablation studies.

Meng: I wonder if this level of structure is something we can actually deploy easily, or if it requires an incredibly complex pipeline just to get started?

Lu: The framework is designed to handle the complexity by formalizing compliance units as four-tuples with a schema, which helps standardize how the rules are represented, making them machine-readable.

Conclusion: Tom: So, looking at "GraphCompliance: Aligning Policy and Context Graphs for LLM-Based Regulatory Compliance," it’s clear the authors are focused on building a system that bridges the gap between rigid legal text and flexible, real-world situations.

Jane: They are essentially showing that combining structured policy representation with contextual knowledge can lead to better performance when AI is tasked with making compliance judgments.

Lu: The implication here is that for anyone working in this space, it points toward a path where neuro-symbolic approaches—using the structure of the law—can be much more effective than purely end-to-end methods for these kinds of tasks.

Meng: For production, it suggests that investing time upfront into building those structured representations might pay off by making the downstream model much more predictable and less likely to produce errors under stress.

Tom: Exactly. It’s about moving from a system that just predicts based on general language to one that grounds its predictions in the actual structure of the regulation, which is what this paper achieves.

Jane: So, it’s about using graph alignment as a way to impose order so that AI reasoning becomes more accurate when dealing with complex, multi-document regulatory environments.

Lu: This gives a solid direction for future work toward extending this framework to other regulatory domains, like finance and healthcare, by validating its policy-agnostic design.

Meng: It seems like the core idea is moving the focus from just being smart in general language to being accurate within a specific domain's rules, which is what this paper addresses.

Tom: Yeah. The paper gives us a solid foundation for building more reliable systems for normative reasoning by using this structure-aware framework.

Seoul National University · Osaka University

cs.AI, cs.IR

Submitted: 2025-10-30

Updated: 2026-10-08

Comments: We acknowledge Makoto Onizuka and Sungmok Kim's contributions to v1. This version substantially reformulates the research question, assumptions, and conceptual framing, with revised methodology and evaluation. Following a reassessment of contributions to this version, authorship was revised with written consent from all v1 authors, including those no longer listed. Submitted to IEEE Access

Code: https://github.com/CHUNG-JiSeong/GraphCompliance

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: The gist: GraphCompliance introduces a framework that represents regulatory texts as a Policy Graph and runtime contexts as a Context Graph, and aligns them to anchor the reasoning of large language

Key concepts

Policy Graph: This graph represents regulatory documents by formalizing them into premises (definitions) and compliance units (rules). Each rule is structured as a 4-tuple, making the logical structure of the regulation explicit. This helps capture how obligations are defined and applied.
Context Graph: The Context Graph formalizes situational facts extracted from unstructured text. It consists of entity–relation triples that describe what is happening in a specific case, mapping real-world events to policy-guided concepts. This organizes contextual data into a comparable format.
Compliance Gate: This is the core alignment engine that performs deterministic structural analysis between the Policy Graph and Context Graph. It identifies relevant regulatory requirements by matching factual anchors in the context against the structured rules in the policy graph, simplifying complex problems for the LLM.
Evidence Window: This mechanism is used during judgment to provide a focused set of evidence to an LLM when it makes a compliance decision. It ensures that the model reasons based only on highly relevant, pre-selected regulatory units and contextual facts.

Terminology

Alignment — 2× in this paper
Ablation — 1× in this paper · explained in 1 episode
Ablation Studies — 1× in this paper · explained in 1 episode
Core — 1× in this paper · explained in 2 episodes
Large Language Model (LLM) — 1× in this paper · explained in 3 episodes
RAG — 1× in this paper · explained in 3 episodes

Summary

The gist: GraphCompliance introduces a framework that represents regulatory texts as a Policy Graph and runtime contexts as a Context Graph, and aligns them to anchor the reasoning of large language models in structured information to reduce regulatory interpretation burden.

How it works

GraphCompliance is a framework that constructs two knowledge graphs (KGs) from policy documents and a given context: a policy graph that captures the logical structure of regulations, and a context graph that formalizes the situational facts The framework aligns these KGs by a compliance gate, which performs deterministic structural analysis before presenting a constrained and simplified problem to the LLM. This approach distinguishes it from prior work by using explicit structural lookups—such as traversing cross-references or checking actor attributes—handled by reliable graph traversal rather than the LLM’s general contextual understanding.

The framework consists of three main components:

Policy Graph Construction: This stage converts regulatory text into premises and compliance units (CUs), each formalized as a 4-tuple schema, r = ⟨S, Θ, Π, κ⟩. Premises denote non-deontic definitional or interpretive material such as terms and role definitions. Compliance units encode obligations, prohibitions, or permissions addressed to a role-bearing actor and constitute the units actually judged in a case.
Context Graph Construction: This stage extracts entity–relation (ER) triples from the unstructured text and maps entities to policy-guided hypernyms. The process involves extracting ER triples using an LLM-based method and then performing hypernym mapping by retrieving top-M policy fragments to guide the mapping.
Compliance Gate Reasoning: This is the core reasoning engine that takes the constructed policy graph (GP) and context graph (GC) as inputs to produce a final compliance judgment. The process starts by retrieving candidate CUs from GP for each factual anchor in GC using a three-part biencoder score. This is followed by cross-encoder reranking to produce a concise and highly relevant CU Plan. The LLM then performs judgment on this plan using an evidence window. A crucial post-processing step involves computing the reference closure R(c) for any initially deemed NON COMPLIANT CU to determine if any CU within R(c) constitutes a valid exception that overrides the initial violation.

Key Contributions and Performance

The paper demonstrates superior performance over baselines, yielding 4.1–7.2 percentage points (pp) higher micro-F1 than the case of LLM-only and RAG baselines in experiments on 300 GDPR-derived real-world scenarios. The framework shows its largest gains on the F2, which weights recall twice as much as precision, indicating a substantially higher recall priority. Ablation studies confirm that structured representations and a judge LLM are complementary for normative reasoning.

Experimental Validation

The framework was evaluated on the GCS-300 benchmark, a semi-synthetic dataset of 300 scenarios grounded in publicly documented enforcement decisions. The evaluation protocol uses macro-F1 and micro-F1 metrics, with F2 score used to reflect recall priority. Results show that GraphCompliance achieves clear improvements across all models compared to a strong RAG baseline. The largest drop in performance occurs when the Context Graph is replaced with raw text, indicating that the subgraph-based anchoring effect is critical for performance.

Submodule Fidelity and Robustness

The fidelity of the intermediate representations is validated through reconstruction tests and cycle-consistency checks. Both the policy graph and context graph exhibit high cycle-consistency, with semantic similarity scores remaining stable after five cycles. Furthermore, noise injection tests show that corrupting the policy graph leads to a measurable drop in semantic similarity, indicating its robustness. The framework is also highly robust to prompt variations, with the F1 range being low at 6.5 pp for GraphCompliance.

Conclusion and Outlook

GraphCompliance proposes a hybrid framework that converts policies and contexts into a Policy Graph and a Context Graph, then aligns them via a Compliance Gate to structurally guide the final LLM-based judgment. The work suggests that this structured, neuro-symbolic approach significantly improves accuracy, robustness, and fidelity over standard end-to-end baselines. Future work is directed toward extending the framework to broader regulatory domains such as finance and healthcare to validate its policy-agnostic design. The paper provides a strong foundation for future research into reliable and verifiable neuro-symbolic systems for normative reasoning.

Improvements for AI systems

Bold header: Policy Graph Construction via LLM-assisted Classification

This improves system accuracy by formalizing regulatory text into structured knowledge graphs, as detailed in Algorithm 1 BuildPolicyGraph, which converts regulatory text into premises and compliance units (CUs), each formalized as: r = ⟨S, Θ, Π, κ⟩.

Bold header: Context Graph Construction with Policy-Guided Hypernym Mapping

This enhancement stabilizes downstream reasoning by explicitly linking extracted real-world entities to regulatory terms through a process where we use the previously built policy graph as retrieval information to guide the mapping, ensuring context entities are mapped to formal policy hypernyms.

Bold header: Compliance Gate using Deterministic Structural Analysis

The system can perform explicit structural lookups—such as traversing cross-references or checking actor attributes—are handled by reliable graph traversal rather than the LLM’s general contextual understanding, which ensures compliance decisions are grounded in deterministic structural analysis before semantic judgment.

Bold header: Defeasible Logic Implementation via Reference Closure

The improved system can handle exceptions robustly by implementing a mechanism where a second LLM call then determines whether any CU within this R(c) constitutes a valid exception that overrides the initial violation, leading to the post-override judgment: yˆ′c = (COMPLIANT, if ∃r ∈ R(c) s.t. IsException(r,W (a)) = true, yˆc, otherwise.

Bold header: Superior Recall through F2-Score Weighting

The framework can prioritize the detection of severe violations by using the F2-score (beta = 2) as a key metric, which weighs recall twice as heavily as precision, resulting in a system that is less likely to overlook high-risk non-compliance and supports risk-aware triage.