Relation Hallucination Index for Evaluating Abstractive Summarization
August 11, 2026
The Relation Hallucination Index (RHI) introduces a dependency-aware algorithm to detect fabricated entity-event relationships in summaries. The framework uses lemmatization-based normalization and named entity grounded subject resolution to improve the fidelity of extracted relation triples.
HOW THIS AFFECTS YOU
●
builderUse RHI-based evaluation to harden your summarization pipelines against high-stakes factual errors.
●
researcherThis provides a more granular metric for evaluating faithfulness beyond standard ROUGE scores.