Knowledge Graph Validation by Integrating LLMs and Human-in-the-Loop#

Original

Authors: stefani-tsaneva, danilo-dessi, francesco-osborne, marta-sabou
Published: Information Processing & Management (2025)
DOI: 10.1016/j.ipm.2025.104145

Empirical investigation of nine knowledge-graph-validation workflows combining LLMs and human-in-the-loop methods, tested on the Computer Science Knowledge Graph (CS-KG, 3.6K triples from 6.7M publications). The paper directly addresses which automation levels balance validation precision/recall while minimizing human effort.

Core Finding: The Disagreement Strategy#

Workflow 5 & 6 (Human Verification): When two automated validators disagree, a human decides. Achieves:

F1 score: 82% (+5% vs baseline SCICERO)
Precision: 80%+
Human annotations: <13% of total triples
No precision/recall tradeoff

This pattern resolves the classic tension: pure automation leaves orphaned triples unchecked, pure manual doesn’t scale, but “validate only on disagreement” captures the best of both.

Nine Workflows Across Four Automation Levels#

1. Human Judgment (Workflows 1-3)#

Workflow 1: Pure manual (no automated support). High precision (+18%), poor scalability (all triples reviewed).
Workflow 2: Human validation after SCICERO automated filters. Precision +8-18%, F1 +4% (1,800+ triples to review).
Workflow 3: Partial human for low-support triples only. Precision +8%, F1 +3% (fewer manual annotations).

Tradeoff: Precision gains come with recall losses (-8%), labor-intensive.

2. AI Assistance (Workflow 4)#

Balanced task partitioning: LLMs validate high-support triples, humans validate low-support triples.

Precision: +15% (up to 90% with GPT-4o)
Recall: -13% (significant loss)
Verdict: Good for precision-critical KGs (medical), poor for coverage-critical KGs.

3. Human Verification (Workflows 5-6)#

Workflow 5: Human validation upon disagreement (both add/remove decisions).

Precision +5%, F1 +2-3%
<13% human effort, balanced improvements

Workflow 6: Human validation only on removal disagreement (coverage prioritized).

Recall +6-8%
Same low human effort
Use case: When KG coverage is priority (research, exploration).

4. Fully Automated (Workflows 7-9)#

Workflow 7: LLM validator as final stage (Precision +12%, F1 +5-7% with GPT-4o)
Workflow 8: LLM for uncertain triples only (Precision +3-8%, no human effort)
Workflow 9: LLM replaces SCICERO entirely (Precision -5%, F1 -3%) — weakest approach alone

Key insight: LLMs integrated with other methods (7/8) outperform LLMs standalone (9). Consensus validation > single validator.

Two Experiments#

Experiment A: Large-Scale Simulation (CS-KG-3600)#

3,600 triples across six categories (high support, low support, transformer-rejected, ontology-rejected, random corruptions). Human validation simulated using expert annotations from original CS-KG evaluation.

Findings:

LLM-based validation increases precision 3-12% without manual effort (workflows 7/8)
Human judgment improves precision 8-18% but requires 1,800+ reviews (workflows 1/2)
Hybrid approaches (workflows 4-6) balance performance and effort

Experiment B: Real-Life Validation (CS-KG-600)#

Subset of 600 triples validated by four PhD researchers in Computer Science (not original evaluators). Only 333 triples sent to human validators (workflows 2-9, skipping pure manual).

Findings:

Fully automated workflows show +5-12% precision improvement (GPT-4o)
Human judgment workflows show smaller improvements (+4-8%) than Experiment A (junior vs. senior experts)
Workflow 8 with Claude Sonnet achieves balanced improvements across all metrics
Key difference: Real humans underperform simulated ground truth, but LLMs remain consistent

Technical Implementation Details#

LLM validators tested:

GPT-4o (gpt-4o-2024-05-13)
Claude Sonnet (claude-3-5-sonnet-20241022)
Llama 3.3 70B (open-source alternative)

Prompting strategy:

Binary task: classify triple as true/false
Batch size: 100 triples per request
Aggregation: majority vote across 3 runs per batch
Context: No external sources, pure LLM reasoning

KG generation pipeline (SCICERO):

Extraction: CSO classifier + NLP modules → candidate triples
Entity/relationship handling: merge, filter generic terms
Validation: transformer validator (support-level based) + ontology validator (schema-based)

Metrics & Trade-offs#

Workflow	Type	Precision	Recall	F1	Human Effort	LLM Calls
SCICERO baseline	—	75%	79%	77%	0	0
Workflow 5	Hybrid	80%	77%	78%	13% of triples	100%
Workflow 6	Hybrid	78%	83%	80%	<13%	100%
Workflow 7	Auto	87%	74%	80%	0	100%
Workflow 9	Auto	70%	75%	72%	0	100%

Strongest paper claim: “Both fully manual and fully automated validation approaches present trade-offs between precision and recall; a hybrid approach, leveraging human-in-the-loop only upon disagreement among automated methods, leads to highest F1 score (+5%) with minimal manual efforts.”

Connections to Wiki#

This paper validates three wiki themes empirically:

human-in-the-loop: Formalized as spectrum from pure manual → hybrid → pure auto. Disagreement strategy is concrete UX pattern matching agentic-ux-patterns.
skill-evaluation: KG validation is a concrete instance of three-tier skill evaluation (transformer baseline → LLM validator → human verification). Workflows 5/6 map to “human spot-checks disagreement” tier.
agent-memory-persistence: KGs are structured memory. Validation quality directly impacts downstream agent reliability — connects to memory-lifecycle-drift (confidence scoring).

Limitations & Open Questions#

Single KG domain: CS-KG designed for Computer Science. Adaptability to medical, legal, e-commerce domains unclear (but workflows domain-independent by design).
LLM variability: Different prompts, model versions, temperature settings not explored. RAG-enhanced validation not tested.
Scalability remaining challenge: Hybrid workflows reduce manual effort but don’t fully solve millions-of-triples KGs. Future work: dynamic workflow selection based on real-time resource availability.
Ground truth validity: LLMs sometimes outperform junior human experts — raises question: which defines truth?

Kg Validation Hybrid Workflows