UCVE: Graph Structured Retrieval for Compliance Validation

Emily Chen; Michelle Menachery; Himanshu Kumar; Rahul Nenavath; Amirfarrokh Iranitalab

cs.CLOct 27, 2026

UCVE: Graph Structured Retrieval for Compliance Validation

Authors: Emily Chen, Michelle Menachery, Himanshu Kumar, Rahul Nenavath, Amirfarrokh Iranitalab

Organizations: Capital One, AI Foundations

Abstract

Structured compliance validation requires determining whether every obligation in a predefined requirement space is satisfied, violated, or unaddressed by a given document. This can be described as a coverage problem that existing retrieval and classification systems, by treating requirements as independent, are not designed to solve. We introduce \ac{UCVE}, a domain-agnostic framework that represents compliance requirement spaces as a \ac{CRG} and performs retrieval via spectral reranking over its reinforcement learning-weighted Laplacian. Edges between requirement nodes are learned via reinforcement learning with a traversal reward grounded in violation detection, reshaping the graph's community structure such that low-pass spectral filtering distinguishes relevant from irrelevant obligation clusters. Evaluating across three compliance domains, \ac{UCVE} averages 4.3% matched- $K$ recall lift over flat cosine retrieval and consistently outperforms strong retrieval and graph-based baselines.

Figures & tables

Figure 1: Universal content validation engine architecture.

	OPP-115			CLAUDETTE			ESRS
Method	Recall	Precision	$K$	Recall	Precision	$K$	Recall	Precision	$K$
BM25	0.649	0.163	14.0	0.732	0.104	16.0	0.4902	0.0640	16.9
Flat cosine	0.748	0.187	14.0	0.806	0.112	16.0	0.7227	0.0967	16.9
Cross-encoder	0.781	0.197	14.0	0.825	0.113	16.0	0.465	0.0590	16.9
PPR	0.583	0.201	14.0	0.731	0.106	16.0	0.5919	0.0825	16.9
Leiden	0.766	0.169	14.0	0.841	0.119	16.0	0.6745	0.0908	16.9

Table 1: Retrieval results at matched-

K

(OPP-115

K{=}14

, CLAUDETTE

K{=}16

, ESRS

K{=}17

). PPR = Personalized PageRank; Leiden = GraphRAG community detection.

Edge Type	OPP-115	CLA	ESRS
cosine_sim	1.000	1.000	1.000
same_category	0.549	N/A	0.500
same_sub_attribute	0.556	N/A	N/A
same_fairness_tier	N/A	0.037	N/A
same_kb_tag	N/A	0.041	N/A
same_parent_cat	N/A	0.040	N/A

Table 2: Learned edge weights per dataset.

Dataset	Strategy	Prec	Rec	$\Delta$ Rec $\uparrow$ †
OPP-115	Retrieval ceiling	—	.792	—
	Baseline	.871	.535	$-$ .257
	Persona	.876	.543	$-$ .249
CLA.	Retrieval ceiling	—	.860	—
	Baseline	.730	.817	$-$ .043
	Persona	.732	.803	$-$ .057

Table 3: End-to-end validation performance with graph-derived context from the CRG using Llama-3.3-70B.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Edge Type Removed	OPP-115 $\Delta$ R	CLAUDETTE $\Delta$ R	ESRS $\Delta$ R
category	-0.027	+0.000	-0.026
fairness_tier	+0.000	+0.000	+0.000
kb_tag	+0.000	+0.000	+0.000
parent_category	+0.000	+0.000	+0.000
sub_attribute	-0.006	+0.000	+0.000
framework	+0.000	+0.000	+0.000

Appendix

Table 4: Recall change when removing individual metadata edge types.

Dataset	Condition	Recall	Precision
OPP	Flat cosine	0.748	0.195
	Spectral $\times$ Cos.	0.759	0.170
	BFS $\times$ RL (@ 14)	0.710	0.184
	Spectral $\times$ RL	0.792	0.185
CLA	Flat cosine	0.806	0.122
	Spectral $\times$ Cos.	0.848	0.117

Appendix

Table 5: Component analysis of graph type

\times

retrieval method. Each condition uses matched-

K

while spectral conditions use natural stopping.

ID	Size	Label	Categories
C1	7	Core Practices	data_retention, data_security, 1st_party_collection, other, 3rd_party_sharing, user_access, user_choice
C2	10	Collection	First Party Collection/Use sub-attributes (9), Do Not Track (1)
C3	2	Jurisdictional	International and Specific Audiences (2)
C4	2	Policy Meta	Policy Change (1), Other (1)
C5	3	Third-Party	Third Party Sharing/Collection sub-attributes (3)

Appendix

Table 6: OPP-115 Spectral Communities (

k=5,k\text{-NN}=3,n=24

ID	Size	Label	Categories (Top-1)
C1	9	Content Removal	CR (9) — 100%
C2	17	Liability Limit.	LTD (17) — 100%
C3	5	Arbitration	A (5) — 100%
C4	15	Termination	TER (15) — 100%
C5	7	Unilat. Change	CH (7) — 100%

Appendix

Table 7: CLAUDETTE Spectral Communities (

k=5,k\text{-NN}=3,n=53

ID	Sz.	Label	Categories	Pillar
C1	12	Gov. & Strat.	Governance (5), Strategy (5), General (2)	G
C2	7	Workforce	Own Workforce (6), Business Conduct (1)	S
C3	2	Anti-corrupt.	Business Conduct (2)	G
C4	26	Environmental	Climate (9), Pollution (6), Circ. Economy (6), Water (5)	E
C5	3	Ethics	Business Conduct (3)	G
C6	11	Labor Stds.	Own Workforce (11)	S

Appendix

Table 8: ESRS-82 Spectral Communities (

k=8,k\text{-NN}=5,n=82

Figure 2: Community structure of the Contextual Requirement Graph across three compliance domains. (a–c) Dense cosine-similarity graphs (57–100% edge density) where near-complete connectivity obscures meaningful topology. (d–f) After edge weight learning and k-NN sparsification, distinct spectral communities emerge (dashed hulls; 5 eigenvectors for OPP-115/CLAUDETTE, 12 for ESRS-82). Node colors indicate requirement categories. This learned sparse structure is the signal spectral reranking exploits for +3.4–4.5% recall lift over flat cosine retrieval.

Figure 3: Spectral reranking hyperparameter sensitivity. Each panel sweeps one parameter while holding the others fixed at their chosen values ( $k_{\text{NN}}=3$ , $k_{\text{eig}}=5$ , $\alpha=0.8$ ). Stars mark chosen values; dotted lines show flat cosine baselines. Performance is stable across wide ranges for $k_{\text{eig}}$ and $\alpha$ , but degrades sharply at $k_{\text{NN}}=1$ where over-sparsification disconnects the graph.

Dataset	Entry Strategy	Rec.	Eff.	C70	Vis.
OPP-115	Random	0.967	0.160	0.725	22.6
	Cosine	0.847	0.177	0.646	17.6
	GSAEPS	0.865	0.174	0.657	18.4
CLA.	Random	0.998	0.051	0.713	47.1
	Cosine	0.967	0.078	0.700	30.5
	GSAEPS	0.967	0.078	0.700	30.6

Appendix

Table 9: Effect of entry point strategy on downstream traversal recall.

Feature	Description
Query Sim	Cosine similarity between $e_{q}$ and $e_{i}$
Neighbor Avg Wght	Mean edge weight from $v_{i}$ to 5 most similar nodes
Parent Avg Sim	Mean query similarity of same-parent nodes
Parent Max Sim	Max query similarity among same-parent nodes
Same Parent Top-1	1 if $v_{i}$ shares parent category with top node
Query Z-Score	Z-scored query similarity within dataset

Appendix

Table 10: Features used in graph-enhanced node scoring.

Method	Encoder	OPP-115 R	CLAUDETTE R	ESRS-82 R	Avg R
BM25	mpnet	0.649	0.732	0.490	0.624
Flat Cosine	mpnet	0.748	0.806	0.720	0.758
PPR	mpnet	0.583	0.731	0.592	0.635
Leiden	mpnet	0.766	0.841	0.675	0.761
Heat Kernel	mpnet	0.747	0.817	0.744	0.769
CRG + Spectral	mpnet	0.792	0.860	0.755	0.802

Appendix

Table 11: Model Performance Evaluation Across Encoders

Data	Model	Acc	F1	Y-P	Y-R	Y-F1
OPP	70B	0.396	0.392	0.948	0.288	0.442
CLA	8B	0.826	0.559	0.121	0.974	0.215
ESRS	8B	0.870	0.571	0.130	0.604	0.213

Appendix

Table 12: Graph-document alignment baseline. Y-P and Y-R denote precision/recall over the positive class.

Method	Rec.@K	$K_{\text{match}}$	Clamp	Tuned	$\Delta$
OPP-115 (target $K$ =14.19)
Flat cosine	.749	15.63	.696	.746	$-.037$ [ $-$ .057, $-$ .016]
Heat kernel	.754	15.41	.795	.749	$-.058$ [ $-$ .073, $-$ .042]
Cross-encoder	.780	14.48	.714	.774	$-.024$ [ $-$ .046, $-$ .004]
CRG + Spectral	.791	–	.791	.756	–
CLAUDETTE (target $K$ =16.42)

Appendix

Table 13: Adaptive stopping robustness check. Rec.@K: recall at CRG + Spectral’s matched-K budget.

K_{\text{match}}

: average K the baseline needs to reach CRG + Spectral’s own recall. Clamp: recall under the unmodified production stopping rule, no calibration. Tuned: recall at the independently calibrated operating point.

\Delta

: bootstrap 95% CI (OPP-115, CLAUDETTE) or mean difference across folds (ESRS), baseline minus CRG + Spectral at matched K.

Dataset	Strategy	Prec	Rec	$\Delta$ Rec $\uparrow$ †
OPP-115	Retrieval ceiling	—	.792	—
	Baseline	.775	.564	$-$ .228
	Persona	.782	.567	$-$ .225
CLA.	Retrieval ceiling	—	.860	—
	Baseline	.748	.697	$-$ .163
	Persona	.773	.685	$-$ .175

Appendix

Table 14: End-to-end validation performance with graph-derived context from the CRG using Llama-3.1-8B.

Pillar	ESG	P	R	P	R	Supp.
		70B		8B
general	G	0.593	0.421	0.269	0.474	38
business conduct	G	0.458	0.491	0.410	0.472	54
water marine	E	0.467	0.250	0.315	0.304	56
consumers / end users	S	0.434	0.434	0.186	0.491	53
value chain workers	S	0.405	0.464	0.218	0.603	68

Appendix

Table 15: Per-pillar end-to-end precision (P) and recall (R) on ESRS under the baseline validation strategy, for Llama-3.3-70B and Llama-3.1-8B . Support is the approximate number of positive instances per pillar. E/S/G indicates the ESRS pillar category (Environmental, Social, Governance).

Graph-Based Retrieval

Graph-based retrieval uses links among documents, entities, or vectors to find connected evidence beyond isolated similarity matches.

See all 142 papers

Citation-Closure Retrieval and Per-Rule Attribution for Real-World Regulatory Compliance Question Answering
May 28, 2026Yeong-Joon Ju, Seong-Whan LeeGraph Retrieval-Augmented Generation Legal QA
Department of Artificial Intelligence, Korea University
The Violation Situation Pattern: A Knowledge-Graph Pattern for Compliance Violations
Jun 2, 2026Nima Kamali Lassem, Fuqi Song, Seyid Amjad AliKnowledge Graphs
DiliTrust, Paris, France · Department of Information Systems and Technologies, Bilkent University, Türkiye
ComplianceNLP: Knowledge-Graph-Augmented RAG for Multi-Framework Regulatory Gap Detection
Apr 26, 2026Dongxin Guo, Jikun Wu, Siu Ming YiuFinancial Services Graph Retrieval-Augmented Generation
The University of Hong Kong · Stellaris AI Limited

May 28, 2026J/K move · Enter open · S save

Citation-Closure Retrieval and Per-Rule Attribution for Real-World Regulatory Compliance Question Answering

Yeong-Joon Ju, Seong-Whan Lee

Department of Artificial Intelligence, Korea University

Graph Retrieval-Augmented Generation Legal QA Evidence Retrieval AI Regulatory Compliance

Open paper AskPDF arXiv

Deploying Large Language Models (LLMs) for regulatory compliance demands rigorous traceability via comprehensive citations across multi-tiered authority structures. Unlike traditional multi-hop or legal QA, this task requires structured procedural lookups and evidence-set closure rather than entity resolution or case-law reasoning. Existing RAG systems struggle here due to flattened citation edges, fragmented retrieval expansions, and fragile post-hoc attribution. We formalize Regulatory Compliance QA with RegOps-Bench, a novel benchmark featuring an Operational Knowledge Graph derived from complex national R&D regulations. To address these bottlenecks, we propose RefWalk, a unified framework driven by a shared topic anchor. RefWalk traverses cross-document citations, fuses multi-view candidates via max-based aggregation, and enforces per-rule attribution to explicitly map claims to sources. We establish a strong baseline with substantial improvements in retrieval recall and citation accuracy. Finally, a contrastive evaluation on a U.S. health compliance dataset (HIPAA) reveals that existing systems exhibit saturation on flat-structure rules, underscoring the need for RegOps-Bench. Our code is available at https://github.com/yeongjoonJu/RefWalk.

Abstract

Graph-Based Retrieval

Explore similar work

Citation-Closure Retrieval and Per-Rule Attribution for Real-World Regulatory Compliance Question Answering

The Violation Situation Pattern: A Knowledge-Graph Pattern for Compliance Violations

ComplianceNLP: Knowledge-Graph-Augmented RAG for Multi-Framework Regulatory Gap Detection