Research map / Large language models / Language, safety and interpretability

Enhancing Uncertainty Modeling with Semantic Graph for Hallucination Detection

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes semantic-graph (AMR-based) uncertainty modeling for LLM hallucination detection, with relation-based uncertainty propagation at sentence level and NLI-based graph calibration at passage level, evaluated on WikiBio and a new Chinese NoteSum dataset.

Problem

Uncertainty-based hallucination detection treats tokens independently and averages sentence scores, missing entity dependencies and contradictions between sentences that span multiple tokens and sentences.

Approach

Build a passage-level AMR semantic graph linking entities and sentences; propagate uncertainty from subject to object entities along graph relations with a relation-intensity penalty; and calibrate passage uncertainty with NLI contradiction probabilities between neighboring sentences in the graph.

Key results

• On WikiBio with LLaMA-30B, the method reaches 77.60 Pearson and 74.44 Spearman at passage level, versus 77.15/73.24 for FOCUS (LLaMA-30B).
• Passage-level Spearman on NoteSum improves to 61.05 versus 40.09 for the best FOCUS baseline, the reported 19.78% maximum gain.
• Ablation: removing graph-based calibration lowers WikiBio passage Pearson from 77.60 to 75.89.

Measured effects reported in the paper

Surprising findings

• Removing the maximum probability term causes passage-level Pearson to collapse from 77.60 to 23.32 on WikiBio.
• Averaging sentence uncertainties and adjacent-only calibration perform similarly, indicating limits of local-only relation modeling.

Limitations

• NoteSum is a private industrial dataset and is small (200 passages), limiting reproducibility.
• Relies on AMR parsing, coreference, and entity linking pipelines whose errors propagate into the graph.
• Requires access to token probabilities and attention scores from the LLM, so it does not apply to black-box models.

Experimental setup

Datasets: WikiBio NoteSum

Benchmarks: WikiBio NoteSum

Metrics: AUC Pearson correlation Spearman correlation

Baselines: GPT-3 Uncertainties SelfCheckGPT FOCUS

Builds on: FOCUS SelfCheckGPT

Closest papers by meaning

• Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling (AAAI 2025 · similarity 0.92)
• Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.91)
• Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.90)
• HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.90)
• Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.89)
• LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.89)