HalluGraph: Auditable Hallucination Detection for Legal RAG Systems via Knowledge Graph Alignment

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Noël, Valentin, Seidou, Elimane Yassine, Capo-Chichi, Charly Ken, Amari, Ghanem
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866908683995185152
author Noël, Valentin
Seidou, Elimane Yassine
Capo-Chichi, Charly Ken
Amari, Ghanem
author_facet Noël, Valentin
Seidou, Elimane Yassine
Capo-Chichi, Charly Ken
Amari, Ghanem
contents Legal AI systems powered by retrieval-augmented generation (RAG) face a critical accountability challenge: when an AI assistant cites case law, statutes, or contractual clauses, practitioners need verifiable guarantees that generated text faithfully represents source documents. Existing hallucination detectors rely on semantic similarity metrics that tolerate entity substitutions, a dangerous failure mode when confusing parties, dates, or legal provisions can have material consequences. We introduce HalluGraph, a graph-theoretic framework that quantifies hallucinations through structural alignment between knowledge graphs extracted from context, query, and response. Our approach produces bounded, interpretable metrics decomposed into \textit{Entity Grounding} (EG), measuring whether entities in the response appear in source documents, and \textit{Relation Preservation} (RP), verifying that asserted relationships are supported by context. On structured control documents, HalluGraph achieves near-perfect discrimination ($>$400 words, $>$20 entities), HalluGraph achieves $AUC = 0.979$, while maintaining robust performance ($AUC \approx 0.89$) on challenging generative legal task, consistently outperforming semantic similarity baselines. The framework provides the transparency and traceability required for high-stakes legal applications, enabling full audit trails from generated assertions back to source passages.
format Preprint
id arxiv_https___arxiv_org_abs_2512_01659
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle HalluGraph: Auditable Hallucination Detection for Legal RAG Systems via Knowledge Graph Alignment
Noël, Valentin
Seidou, Elimane Yassine
Capo-Chichi, Charly Ken
Amari, Ghanem
Machine Learning
Artificial Intelligence
Computation and Language
Legal AI systems powered by retrieval-augmented generation (RAG) face a critical accountability challenge: when an AI assistant cites case law, statutes, or contractual clauses, practitioners need verifiable guarantees that generated text faithfully represents source documents. Existing hallucination detectors rely on semantic similarity metrics that tolerate entity substitutions, a dangerous failure mode when confusing parties, dates, or legal provisions can have material consequences. We introduce HalluGraph, a graph-theoretic framework that quantifies hallucinations through structural alignment between knowledge graphs extracted from context, query, and response. Our approach produces bounded, interpretable metrics decomposed into \textit{Entity Grounding} (EG), measuring whether entities in the response appear in source documents, and \textit{Relation Preservation} (RP), verifying that asserted relationships are supported by context. On structured control documents, HalluGraph achieves near-perfect discrimination ($>$400 words, $>$20 entities), HalluGraph achieves $AUC = 0.979$, while maintaining robust performance ($AUC \approx 0.89$) on challenging generative legal task, consistently outperforming semantic similarity baselines. The framework provides the transparency and traceability required for high-stakes legal applications, enabling full audit trails from generated assertions back to source passages.
title HalluGraph: Auditable Hallucination Detection for Legal RAG Systems via Knowledge Graph Alignment
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2512.01659