Conceptos Básicos
VERIFINER is a post-hoc verification framework that leverages knowledge and large language models to correct errors in NER predictions.
Resumen
Recent advances in domain-specific NER lack faithfulness, leading to erroneous predictions.
VERIFINER proposes a framework to verify errors from existing NER methods using knowledge and large language models.
The framework validates its effectiveness through experiments on biomedical datasets.
VERIFINER successfully corrects errors without re-training models, showing promise for real-world applications.
The framework consists of span factuality verification, type factuality verification, and contextual relevance verification modules.
VERIFINER demonstrates significant improvements over baselines in precision and F1 scores.
The framework shows robustness in unseen and shifted distribution settings, as well as in low-resource scenarios.
VERIFINER addresses limitations and ethical considerations related to the use of large language models.
Estadísticas
"The results suggest that VERIFINER can successfully verify errors from existing models as a model-agnostic approach."
"The results show that VERIFINER consistently achieves significant improvements over initial predictions on both datasets."
"VERIFINER consistently achieves high precision irrespective of the number of training examples."
Citas
"VERIFINER can successfully verify errors from existing models as a model-agnostic approach."
"VERIFINER consistently achieves significant improvements over initial predictions on both datasets."