Faithful detection of hallucinations in event extraction via causal intervention

Le résumé fourni par la source

The deployment of large language models in event extraction is seriously hindered by structural hallucinations where generated event tuples diverge from the source text due to parametric knowledge. Current detection methods relying on self-reflective prompting and observational consistency fail to identify persistent hallucinated relations caused by parametric dominance. To address this limitation, a causal inference framework called Dual-Perspective Intervention Consistency Detection (DICD) is proposed to shift the paradigm from passive evaluation to active structural intervention. The framework operationalizes textual causal faithfulness as a dynamic dependency that demands causal efficacy through sensitivity to trigger falsification and causal specificity through invariance under entity perturbation. By utilizing atomic decomposition of event tuples and counterfactual context manipulation, the method systematically isolates genuine contextual binding from parametric dominance and stochastic noise. Empirical validation on three newly constructed hallucination detection datasets for event extraction demonstrates that the intervention approach achieves superior performance over observational baselines in diagnosing argument fabrications caused by parametric dominance.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.