Aletheia: A Two-Stage Graph-Based Framework for Hallucination Detection in Abstractive Summarization

Abstractive summarization models have made significant strides in advancing fluency, yet their tendency to generate hallucinations remains a substantial challenge for practical deployment. Effective detection of these hallucinations necessitates three key properties: computational efficiency, interpretability, and robustness under source-summary length asymmetry. However, existing solutions often fail to meet all three requirements. To address these limitations, we introduce Aletheia, a lightweight, two-stage graph-based framework designed for hallucination detection in abstractive summarization. It first employs a lightweight entity screening for unambiguous cases, and then leverages Abstract Meaning Representation (AMR) graphs and a novel summary-normalized concept coverage metric for summarization length discrepancies in complex instances. By deploying three typed detectors, our framework explicitly isolates entity, relation, and semantic confusion errors, providing precise localization and interpretable evidence for each identified hallucination. Empirical evaluations on FRANK highlight the efficacy of the framework, with a PR-AUC of 0.90 on the FRANK dataset.